fix nan

.gitignore

@@ -11,4 +11,5 @@ test/
 *.log
 /data_generation/data/*
 /data_generation/output/*
-test/
+test/
+temp*

Readme.md

@@ -0,0 +1,192 @@
+# Next Best View for Reconstruction
+
+## 1. Setup Environment
+### 1.1 Install Main Project
+```bash
+mkdir nbv_rec
+cd nbv_rec
+git clone https://git.hofee.top/hofee/nbv_reconstruction.git
+```
+### 1.2 Install PytorchBoot
+the environment is based on PytorchBoot, clone and install it from [PytorchBoot](https://git.hofee.top/hofee/PyTorchBoot.git)
+```bash
+git clone https://git.hofee.top/hofee/PyTorchBoot.git
+cd PyTorchBoot
+pip install .
+cd ..
+```
+### 1.3 Install Blender (Optional)
+If you want to render your own dataset as described in [section 2. Render Datasets](#2-render-datasets), you'll need to install Blender version 4.0 from [Blender Release](https://download.blender.org/release/Blender4.0/). Here is an example of installing Blender on Ubuntu:
+```bash
+wget https://download.blender.org/release/Blender4.0/blender-4.0.2-linux-x64.tar.xz
+tar -xvf blender-4.0.2-linux-x64.tar.xz
+```
+If blender is not in your PATH, you can add it by:
+```bash
+export PATH=$PATH:/path/to/blender/blender-4.0.2-linux-x64
+```
+To run the blender script, you need to install the `pyyaml` and `scipy` package into your blender python environment. Run the following command to print the python path of your blender:
+```bash
+./blender -b --python-expr "import sys; print(sys.executable)"
+```
+Then copy the python path `/path/to/blender_python` shown in the output and run the following command to install the packages:
+```bash
+/path/to/blender_python -m pip install pyyaml scipy
+```
+### 1.4 Install Blender Render Script (Optional)
+Clone the script from [nbv_rec_blender_render](https://git.hofee.top/hofee/nbv_rec_blender_render.git) and rename it to `blender`:
+```bash
+git clone https://git.hofee.top/hofee/nbv_rec_blender_render.git
+mv nbv_rec_blender_render blender
+```
+
+### 1.5 Check Dependencies
+Switch to the project root directory and run `pytorch-boot scan` or `ptb scan` to check if all dependencies are installed:
+```bash
+cd nbv_reconstruction
+pytorch-boot scan
+# or
+ptb scan
+```
+If you see project structure information in the output, it means all dependencies are correctly installed. Otherwise, you may need to run `pip install xxx` to install the missing packages.
+
+## 2. Render Datasets (Optional)
+### 2.1 Download Object Mesh Models
+Download the mesh models divided into three parts from: 
+ - [object_meshes_part1.zip](None)
+ - [object_meshes_part2.zip](https://pan.baidu.com/s/1pBPhrFtBwEGp1g4vwsLIxA?pwd=1234)
+ - [object_meshes_part3.zip](https://pan.baidu.com/s/1peE8HqFFL0qNFhM5OC69gA?pwd=1234)
+
+or download the whole dataset from [object_meshes.zip](https://pan.baidu.com/s/1ilWWgzg_l7_pPBv64eSgzA?pwd=1234)
+
+Download the table model from [table.obj](https://pan.baidu.com/s/1sjjiID25Es_kmcdUIjU_Dw?pwd=1234)
+
+### 2.2 Set Render Configurations
+Open file `configs/local/view_generate_config.yaml` and modify the parameters to fit your needs. You are required to at least set the following parameters in `runner-generate`:
+ - `object_dir`: the directory of the downloaded object mesh models
+ - `output_dir`: the directory to save the rendered dataset
+ - `table_model_path`: the path of the downloaded table model
+
+### 2.3 Render Dataset
+
+There are two ways to render the dataset:
+
+#### 2.3.1 Render with Visual Monitoring
+
+If you want to visually monitor the rendering progress and machine resource usage:
+
+1. In the terminal, run:
+   ```
+   ptb ui
+   ```
+2. Open your browser and visit http://localhost:5000
+3. Navigate to `Project Dashboard - Project Structure - Applications - generate_view`
+4. Click the `Run` button to execute the rendering script
+
+#### 2.3.2 Render in Terminal
+
+If you don't need visual monitoring and prefer to run the rendering process directly in the terminal, simply run:
+
+```
+ptb run generate_view
+```
+
+This command will start the rendering process without launching the UI.
+
+## 3. Preprocess
+
+⚠️ The preprocessing code is currently not managed by `PytorchBoot`. To run the preprocessing:
+
+1. Open the `./preprocess/preprocessor.py` file.
+2. Locate the `if __name__ == "__main__":` block at the bottom of the file.
+3. Specify the dataset folder by setting `root = "path/to/your/dataset"`.
+4. Run the preprocessing script directly:
+
+   ```
+   python ./preprocess/preprocessor.py
+   ```
+
+This will preprocess the data in the specified dataset folder.
+
+## 4. Generate Strategy Label
+
+### 4.1 Set Configuration
+
+Open the file `configs/local/strategy_generate_config.yaml` and modify the parameters to fit your needs. You are required to at least set the following parameter:
+
+- `datasets.OmniObject3d.root_dir`: the directory of your dataset
+
+### 4.2 Generate Strategy Label
+
+There are two ways to generate the strategy label:
+
+#### 4.2.1 Generate with Visual Monitoring
+
+If you want to visually monitor the generation progress and machine resource usage:
+
+1. In the terminal, run:
+   ```
+   ptb ui
+   ```
+2. Open your browser and visit http://localhost:5000
+3. Navigate to Project Dashboard - Project Structure - Applications - generate_strategy
+4. Click the `Run` button to execute the generation script
+
+#### 4.2.2 Generate in Terminal
+
+If you don't need visual monitoring and prefer to run the generation process directly in the terminal, simply run:
+
+```
+ptb run generate_strategy
+```
+
+This command will start the strategy label generation process without launching the UI.
+
+## 5. Train
+
+### 5.1 Set Configuration
+
+Open the file `configs/local/train_config.yaml` and modify the parameters to fit your needs. You are required to at least set the following parameters in the `experiment` section:
+
+```yaml
+experiment:
+  name: your_experiment_name
+  root_dir: path/to/your/experiment_dir
+  use_checkpoint: False # if True, the checkpoint will be loaded
+  epoch: 600 # specific epoch to load, -1 stands for last epoch
+  max_epochs: 5000 # maximum epochs to train
+  save_checkpoint_interval: 1 # save checkpoint interval
+  test_first: True # if True, test process will be performed before training at each epoch
+```
+
+Adjust these parameters according to your training requirements.
+
+
+### 5.2 Start Training
+
+There are two ways to start the training process:
+
+#### 5.2.1 Train with Visual Monitoring
+
+If you want to visually monitor the training progress and machine resource usage:
+
+1. In the terminal, run:
+   ```
+   ptb ui
+   ```
+2. Open your browser and visit http://localhost:5000
+3. Navigate to Project Dashboard - Project Structure - Applications - train
+4. Click the `Run` button to start the training process
+
+#### 5.2.2 Train in Terminal
+
+If you don't need visual monitoring and prefer to run the training process directly in the terminal, simply run:
+
+```
+ptb run train
+```
+
+This command will start the training process without launching the UI.
+
+## 6. Evaluation
+...

app_split.py

@@ -5,5 +5,5 @@ from runners.data_spliter import DataSpliter
 class DataSplitApp:
     @staticmethod
     def start():
-        DataSpliter("configs/server/split_dataset_config.yaml").run()
+        DataSpliter("configs/server/server_split_dataset_config.yaml").run()
         

app_train.py

@@ -5,4 +5,4 @@ from PytorchBoot.runners.trainer import DefaultTrainer
 class TrainApp:
     @staticmethod
     def start():
-        DefaultTrainer("configs/server/train_config.yaml").run()
+        DefaultTrainer("configs/server/server_train_config.yaml").run()

configs/local/inference_config.yaml

@@ -6,24 +6,24 @@ runner:
     cuda_visible_devices: "0,1,2,3,4,5,6,7"
     
   experiment:
-    name: local_full_eval
+    name: w_gf_wo_lf_full
     root_dir: "experiments"
-    epoch: 20 # -1 stands for last epoch
+    epoch: 1 # -1 stands for last epoch
 
   test:
     dataset_list:
       - OmniObject3d_train
 
   blender_script_path: "/media/hofee/data/project/python/nbv_reconstruction/blender/data_renderer.py"
-  output_dir: "/media/hofee/data/project/python/nbv_reconstruction/sample_for_training/inference_result_full"
-  pipeline: nbv_reconstruction_pipeline
+  output_dir: "/media/hofee/data/project/python/nbv_reconstruction/nbv_reconstruction/test/inference_global_full_on_testset"
+  pipeline: nbv_reconstruction_global_pts_pipeline
   
 dataset:
   OmniObject3d_train:
     root_dir: "/media/hofee/repository/nbv_reconstruction_data_512"
     model_dir: "/media/hofee/data/data/scaled_object_meshes"
     source: seq_nbv_reconstruction_dataset
-    split_file: "/media/hofee/data/project/python/nbv_reconstruction/sample_for_training/OmniObject3d_train.txt"
+    split_file: "/media/hofee/data/project/python/nbv_reconstruction/nbv_reconstruction/test/test_set_list.txt"
     type: test
     filter_degree: 75
     ratio: 1
@@ -33,11 +33,25 @@ dataset:
     load_from_preprocess: False
 
 pipeline:
-  nbv_reconstruction_pipeline:
-    pts_encoder: pointnet_encoder
-    seq_encoder: transformer_seq_encoder
-    pose_encoder: pose_encoder
-    view_finder: gf_view_finder
+  nbv_reconstruction_local_pts_pipeline:
+    modules:
+      pts_encoder: pointnet_encoder
+      seq_encoder: transformer_seq_encoder
+      pose_encoder: pose_encoder
+      view_finder: gf_view_finder
+    eps: 1e-5
+    global_scanned_feat: False
+
+  nbv_reconstruction_global_pts_pipeline:
+    modules:
+      pts_encoder: pointnet_encoder
+      pose_seq_encoder: transformer_pose_seq_encoder
+      pose_encoder: pose_encoder
+      view_finder: gf_view_finder
+    eps: 1e-5
+    global_scanned_feat: True
+
+
     
 module:
 
@@ -55,6 +69,13 @@ module:
     num_layers: 3
     output_dim: 2048
 
+  transformer_pose_seq_encoder:
+    pose_embed_dim: 256
+    num_heads: 4
+    ffn_dim: 256
+    num_layers: 3
+    output_dim: 1024
+
   gf_view_finder:
     t_feat_dim: 128
     pose_feat_dim: 256

configs/local/strategy_generate_config.yaml

@@ -11,24 +11,17 @@ runner:
     root_dir: "experiments"
       
   generate:
-    voxel_threshold: 0.01
-    overlap_threshold: 0.5
-    filter_degree: 75
-    to_specified_dir: True # if True, output_dir is used, otherwise, root_dir is used
-    save_points: True
-    load_points: True
-    save_best_combined_points: False
-    save_mesh: True
+    voxel_threshold: 0.003
+    overlap_area_threshold: 30
+    compute_with_normal: False
+    scan_points_threshold: 10
     overwrite: False
-    seq_num: 50
+    seq_num: 10
     dataset_list:
       - OmniObject3d
   
 datasets:
     OmniObject3d:
-      #"/media/hofee/data/data/temp_output"
-      root_dir: "/media/hofee/data/data/sample_data/view_data" 
-      model_dir: "/media/hofee/data/data/scaled_object_meshes"
-      #output_dir: "/media/hofee/data/data/label_output"
-
-    
+      root_dir: /data/hofee/nbv_rec_part2_preprocessed
+      from: 155
+      to: 165 # ..-1 means end

configs/local/train_config.yaml

@@ -84,7 +84,7 @@ module:
   gf_view_finder:
     t_feat_dim: 128
     pose_feat_dim: 256
-    main_feat_dim: 2048
+    main_feat_dim: 3072
     regression_head: Rx_Ry_and_T
     pose_mode: rot_matrix
     per_point_feature: False

configs/local/view_generate_config.yaml

@@ -7,32 +7,31 @@ runner:
     name: debug
     root_dir: experiments
   generate:
-    object_dir: /media/hofee/data/data/scaled_object_meshes
-    table_model_path: /media/hofee/data/data/others/table.obj
-    output_dir: /media/hofee/repository/new_nbv_reconstruction_data_512
+    port: 5002
+    from: 600
+    to: -1 # -1 means all
+    object_dir: /media/hofee/data/data/object_meshes_part1
+    table_model_path: "/media/hofee/data/data/others/table.obj"
+    output_dir: /media/hofee/repository/data_part_1
     binocular_vision: true
     plane_size: 10
     max_views: 512
-    min_views: 64
+    min_views: 128
+    random_view_ratio: 0.02
+    min_cam_table_included_degree: 20
     max_diag: 0.7
-    min_diag: 0.1
+    min_diag: 0.01
     random_config:
       display_table:
         min_height: 0.05
         max_height: 0.15
-        min_radius: 0.3
-        max_radius: 0.5
-        min_R: 0.05
-        max_R: 0.3
-        min_G: 0.05
-        max_G: 0.3
-        min_B: 0.05
-        max_B: 0.3
+        min_radius: 0.2
+        max_radius: 0.3
       display_object:
         min_x: 0
-        max_x: 0.03
+        max_x: 0.05
         min_y: 0
-        max_y: 0.03
+        max_y: 0.05
         min_z: 0.01
         max_z: 0.01
         random_rotation_ratio: 0.3
@@ -44,10 +43,10 @@ runner:
         near_plane: 0.01
         far_plane: 5
         fov_vertical: 25
-        resolution: [1280,800]
-        eye_distance: 0.15
+        resolution: [640,400]
+        eye_distance: 0.10
         eye_angle: 25
       Light:
         location: [0,0,3.5]
         orientation: [0,0,0]
-        power: 150
+        power: 150

configs/server/server_inference_config.yaml

@@ -0,0 +1,92 @@
+
+runner:
+  general:
+    seed: 1
+    device: cuda
+    cuda_visible_devices: "0,1,2,3,4,5,6,7"
+    
+  experiment:
+    name: w_gf_wo_lf_full_debug
+    root_dir: "experiments"
+    epoch: 1 # -1 stands for last epoch
+
+  test:
+    dataset_list:
+      - OmniObject3d_train
+
+  blender_script_path: ""
+  output_dir: ""
+  pipeline: nbv_reconstruction_global_pts_pipeline
+  
+dataset:
+  OmniObject3d_train:
+    root_dir: "/home/data/hofee/project/nbv_rec/data/nbv_rec_data_512_preproc_npy"
+    model_dir: "/home/data/hofee/project/nbv_rec/data/scaled_object_meshes" 
+    source: seq_nbv_reconstruction_dataset 
+    split_file: "/home/data/hofee/project/nbv_rec/data/OmniObject3d_train.txt" 
+    type: test
+    filter_degree: 75
+    ratio: 1
+    batch_size: 1
+    num_workers: 12
+    pts_num: 4096
+    load_from_preprocess: True
+
+pipeline:
+  nbv_reconstruction_local_pts_pipeline:
+    modules:
+      pts_encoder: pointnet_encoder
+      seq_encoder: transformer_seq_encoder
+      pose_encoder: pose_encoder
+      view_finder: gf_view_finder
+    eps: 1e-5
+    global_scanned_feat: False
+
+  nbv_reconstruction_global_pts_pipeline:
+    modules:
+      pts_encoder: pointnet_encoder
+      pose_seq_encoder: transformer_pose_seq_encoder
+      pose_encoder: pose_encoder
+      view_finder: gf_view_finder
+    eps: 1e-5
+    global_scanned_feat: True
+
+
+    
+module:
+
+  pointnet_encoder:
+    in_dim: 3
+    out_dim: 1024
+    global_feat: True
+    feature_transform: False
+
+  transformer_seq_encoder:
+    pts_embed_dim: 1024
+    pose_embed_dim: 256
+    num_heads: 4
+    ffn_dim: 256
+    num_layers: 3
+    output_dim: 2048
+
+  transformer_pose_seq_encoder:
+    pose_embed_dim: 256
+    num_heads: 4
+    ffn_dim: 256
+    num_layers: 3
+    output_dim: 1024
+
+  gf_view_finder:
+    t_feat_dim: 128
+    pose_feat_dim: 256
+    main_feat_dim: 2048
+    regression_head: Rx_Ry_and_T
+    pose_mode: rot_matrix
+    per_point_feature: False
+    sample_mode: ode
+    sampling_steps: 500
+    sde_mode: ve
+
+  pose_encoder:
+    pose_dim: 9
+    out_dim: 256

configs/server/server_split_dataset_config.yaml

@@ -0,0 +1,22 @@
+
+runner:
+  general:
+    seed: 0
+    device: cpu
+    cuda_visible_devices: "0,1,2,3,4,5,6,7"
+    
+  experiment:
+    name: server_split_dataset
+    root_dir: "experiments"
+  
+  split: # 
+    root_dir: "/data/hofee/data/new_full_data"
+    type: "unseen_instance" # "unseen_category"
+    datasets:
+      OmniObject3d_train: 
+        path: "/data/hofee/data/new_full_data_list/OmniObject3d_train.txt"
+        ratio: 0.9
+
+      OmniObject3d_test: 
+        path: "/data/hofee/data/new_full_data_list/OmniObject3d_test.txt"
+        ratio: 0.1

configs/server/server_train_config.yaml

@@ -3,17 +3,17 @@ runner:
   general:
     seed: 0
     device: cuda
-    cuda_visible_devices: "0,1,2,3,4,5,6,7"
+    cuda_visible_devices: "1"
     parallel: False
     
   experiment:
-    name: new_test_overfit_to_world_preprocessed
+    name: train_ab_global_and_partial_global
     root_dir: "experiments"
     use_checkpoint: False
     epoch: -1 # -1 stands for last epoch
     max_epochs: 5000
-    save_checkpoint_interval: 3
-    test_first: True  
+    save_checkpoint_interval: 1
+    test_first: False  
 
   train:
     optimizer:
@@ -25,46 +25,68 @@ runner:
   test:
     frequency: 3 # test frequency
     dataset_list:
-      - OmniObject3d_test
+      #- OmniObject3d_test
+      - OmniObject3d_val
 
   pipeline: nbv_reconstruction_pipeline
   
 dataset:
   OmniObject3d_train:
-    root_dir: "../data/sample_for_training_preprocessed/sample_preprocessed_scenes"
+    root_dir: "/data/hofee/data/new_full_data"
     model_dir: "../data/scaled_object_meshes"
     source: nbv_reconstruction_dataset
-    split_file: "../data/sample_for_training_preprocessed/OmniObject3d_train.txt"
+    split_file: "/data/hofee/data/new_full_data_list/OmniObject3d_train.txt"
     type: train
     cache: True
     ratio: 1
-    batch_size: 160
-    num_workers: 16
-    pts_num: 4096
+    batch_size: 80
+    num_workers: 128
+    pts_num: 8192
     load_from_preprocess: True
 
   OmniObject3d_test:
-    root_dir: "../data/sample_for_training_preprocessed/sample_preprocessed_scenes"
+    root_dir: "/data/hofee/data/new_full_data"
     model_dir: "../data/scaled_object_meshes"
     source: nbv_reconstruction_dataset
-    split_file: "../data/sample_for_training_preprocessed/OmniObject3d_train.txt"
+    split_file: "/data/hofee/data/new_full_data_list/OmniObject3d_test.txt"
+    type: test
+    cache: True
+    filter_degree: 75
+    eval_list:
+      - pose_diff
+    ratio: 1
+    batch_size: 80
+    num_workers: 12
+    pts_num: 8192
+    load_from_preprocess: True
+
+  OmniObject3d_val:
+    root_dir: "/data/hofee/data/new_full_data"
+    model_dir: "../data/scaled_object_meshes"
+    source: nbv_reconstruction_dataset
+    split_file: "/data/hofee/data/new_full_data_list/OmniObject3d_train.txt"
     type: test
     cache: True
     filter_degree: 75
     eval_list:
       - pose_diff
     ratio: 0.1
-    batch_size: 1
+    batch_size: 80
     num_workers: 12
-    pts_num: 4096
+    pts_num: 8192
     load_from_preprocess: True
 
+
 pipeline:
   nbv_reconstruction_pipeline:
-    pts_encoder: pointnet_encoder
-    seq_encoder: transformer_seq_encoder
-    pose_encoder: pose_encoder
-    view_finder: gf_view_finder
+    modules:
+      pts_encoder: pointnet_encoder
+      seq_encoder: transformer_seq_encoder
+      pose_encoder: pose_encoder
+      view_finder: gf_view_finder
+    eps: 1e-5
+    global_scanned_feat: True
+
     
 module:
 
@@ -75,12 +97,11 @@ module:
     feature_transform: False
 
   transformer_seq_encoder:
-    pts_embed_dim: 1024
-    pose_embed_dim: 256
+    embed_dim: 320
     num_heads: 4
     ffn_dim: 256
     num_layers: 3
-    output_dim: 2048
+    output_dim: 1024
 
   gf_view_finder:
     t_feat_dim: 128
@@ -97,6 +118,9 @@ module:
     pose_dim: 9
     out_dim: 256
 
+  pts_num_encoder:
+    out_dim: 64
+
 loss_function:
   gf_loss:
 

configs/server/split_dataset_config.yaml

@@ -1,22 +0,0 @@
-
-runner:
-  general:
-    seed: 0
-    device: cpu
-    cuda_visible_devices: "0,1,2,3,4,5,6,7"
-    
-  experiment:
-    name: debug
-    root_dir: "experiments"
-  
-  split:
-    root_dir: "../data/sample_for_training_preprocessed/sample_preprocessed_scenes"
-    type: "unseen_instance" # "unseen_category"
-    datasets:
-      OmniObject3d_train: 
-        path: "../data/sample_for_training_preprocessed/OmniObject3d_train.txt"
-        ratio: 0.9
-
-      OmniObject3d_test: 
-        path: "../data/sample_for_training_preprocessed/OmniObject3d_test.txt"
-        ratio: 0.1

configs/server/strategy_generate_config.yaml

@@ -1,32 +0,0 @@
-
-runner:
-  general:
-    seed: 0
-    device: cpu
-    cuda_visible_devices: "0,1,2,3,4,5,6,7"
-
-    
-  experiment:
-    name: debug
-    root_dir: "experiments"
-      
-  generate:
-    voxel_threshold: 0.01
-    overlap_threshold: 0.5
-    filter_degree: 75
-    to_specified_dir: False # if True, output_dir is used, otherwise, root_dir is used
-    save_points: True
-    save_best_combined_points: True
-    save_mesh: True
-    overwrite: False
-    dataset_list:
-      - OmniObject3d
-  
-datasets:
-    OmniObject3d:
-      #"/media/hofee/data/data/temp_output"
-      root_dir: "../data/sample_for_training/scenes" 
-      model_dir: "../data/scaled_object_meshes"
-      #output_dir: "/media/hofee/data/data/label_output"
-
-    

configs/server/view_generate_config.yaml

@@ -1,53 +0,0 @@
-runner:
-  general:
-    seed: 0
-    device: cpu
-    cuda_visible_devices: 0,1,2,3,4,5,6,7
-  experiment:
-    name: debug
-    root_dir: experiments
-  generate:
-    object_dir: ../data/scaled_object_meshes
-    table_model_path: ../data/others/table.obj
-    output_dir: ../data/nbv_reconstruction_data_512
-    binocular_vision: true
-    plane_size: 10
-    max_views: 512
-    min_views: 64
-    max_diag: 0.7
-    min_diag: 0.1
-    random_config:
-      display_table:
-        min_height: 0.05
-        max_height: 0.15
-        min_radius: 0.3
-        max_radius: 0.5
-        min_R: 0.05
-        max_R: 0.3
-        min_G: 0.05
-        max_G: 0.3
-        min_B: 0.05
-        max_B: 0.3
-      display_object:
-        min_x: 0
-        max_x: 0.03
-        min_y: 0
-        max_y: 0.03
-        min_z: 0.01
-        max_z: 0.01
-        random_rotation_ratio: 0.3
-      random_objects:
-        num: 4
-        cluster: 0.9
-    light_and_camera_config:
-      Camera:
-        near_plane: 0.01
-        far_plane: 5
-        fov_vertical: 25
-        resolution: [1280,800]
-        eye_distance: 0.15
-        eye_angle: 25
-      Light:
-        location: [0,0,3.5]
-        orientation: [0,0,0]
-        power: 150

core/evaluation.py

@@ -29,8 +29,9 @@ class PoseDiff:
             gt_rot_mat = PoseUtil.rotation_6d_to_matrix_tensor_batch(gt_rot_6d)
             pred_rot_mat = PoseUtil.rotation_6d_to_matrix_tensor_batch(pred_rot_6d)
             rotation_angles = PoseUtil.rotation_angle_distance(gt_rot_mat, pred_rot_mat)
+            
             rot_angle_list.extend(list(rotation_angles))
-            trans_dist = torch.norm(gt_trans-pred_trans)
+            trans_dist = torch.norm(gt_trans-pred_trans, dim=1).mean().item()
             trans_dist_list.append(trans_dist)
             
 

core/global_pts_pipeline.py

@@ -0,0 +1,94 @@
+import torch
+from torch import nn
+import PytorchBoot.namespace as namespace
+import PytorchBoot.stereotype as stereotype
+from PytorchBoot.factory.component_factory import ComponentFactory
+from PytorchBoot.utils import Log
+
+
+@stereotype.pipeline("nbv_reconstruction_global_pts_pipeline")
+class NBVReconstructionGlobalPointsPipeline(nn.Module):
+    def __init__(self, config):
+        super(NBVReconstructionGlobalPointsPipeline, self).__init__()
+        self.config = config
+        self.module_config = config["modules"]
+        self.pts_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["pts_encoder"])
+        self.pose_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["pose_encoder"])
+        self.pose_seq_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["pose_seq_encoder"])
+        self.view_finder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["view_finder"])
+        self.eps = float(self.config["eps"])
+        self.enable_global_scanned_feat = self.config["global_scanned_feat"]
+        
+    def forward(self, data):
+        mode = data["mode"]
+        
+        if mode == namespace.Mode.TRAIN:
+            return self.forward_train(data)
+        elif mode == namespace.Mode.TEST:
+            return self.forward_test(data)
+        else:
+            Log.error("Unknown mode: {}".format(mode), True)
+    
+    def pertube_data(self, gt_delta_9d):
+        bs = gt_delta_9d.shape[0]
+        random_t = torch.rand(bs, device=gt_delta_9d.device) * (1. - self.eps) + self.eps
+        random_t = random_t.unsqueeze(-1)
+        mu, std = self.view_finder.marginal_prob(gt_delta_9d, random_t)
+        std = std.view(-1, 1)
+        z = torch.randn_like(gt_delta_9d)
+        perturbed_x = mu + z * std
+        target_score = - z * std / (std ** 2)
+        return perturbed_x, random_t, target_score, std
+    
+    def forward_train(self, data):
+        main_feat = self.get_main_feat(data)
+        ''' get std '''
+        best_to_world_pose_9d_batch = data["best_to_world_pose_9d"]
+        perturbed_x, random_t, target_score, std = self.pertube_data(best_to_world_pose_9d_batch)
+        input_data = {
+            "sampled_pose": perturbed_x,
+            "t": random_t,
+            "main_feat": main_feat,
+        }
+        estimated_score = self.view_finder(input_data)
+        output = {
+            "estimated_score": estimated_score,
+            "target_score": target_score,
+            "std": std
+        }
+        return output
+    
+    def forward_test(self,data):
+        main_feat = self.get_main_feat(data)
+        estimated_delta_rot_9d, in_process_sample = self.view_finder.next_best_view(main_feat)
+        result = {
+            "pred_pose_9d": estimated_delta_rot_9d,
+            "in_process_sample": in_process_sample
+        }
+        return result
+    
+    
+    def get_main_feat(self, data):
+        scanned_n_to_world_pose_9d_batch = data['scanned_n_to_world_pose_9d']
+
+        device = next(self.parameters()).device
+
+        pose_feat_seq_list = []
+        
+        for scanned_n_to_world_pose_9d in scanned_n_to_world_pose_9d_batch:
+            scanned_n_to_world_pose_9d = scanned_n_to_world_pose_9d.to(device)
+            pose_feat_seq_list.append(self.pose_encoder.encode_pose(scanned_n_to_world_pose_9d))
+
+        main_feat = self.pose_seq_encoder.encode_sequence(pose_feat_seq_list)
+        
+        
+        combined_scanned_pts_batch = data['combined_scanned_pts']   
+        global_scanned_feat = self.pts_encoder.encode_points(combined_scanned_pts_batch)
+        main_feat = torch.cat([main_feat, global_scanned_feat], dim=-1)
+        
+        
+        if torch.isnan(main_feat).any():
+            Log.error("nan in main_feat", True)
+        
+        return main_feat
+    

core/local_pts_pipeline.py

@@ -0,0 +1,101 @@
+import torch
+from torch import nn
+import PytorchBoot.namespace as namespace
+import PytorchBoot.stereotype as stereotype
+from PytorchBoot.factory.component_factory import ComponentFactory
+from PytorchBoot.utils import Log
+
+@stereotype.pipeline("nbv_reconstruction_local_pts_pipeline")
+class NBVReconstructionLocalPointsPipeline(nn.Module):
+    def __init__(self, config):
+        super(NBVReconstructionLocalPointsPipeline, self).__init__()
+        self.config = config
+        self.module_config = config["modules"]
+        self.pts_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["pts_encoder"])
+        self.pose_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["pose_encoder"])
+        self.seq_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["seq_encoder"])
+        self.view_finder = ComponentFactory.create(namespace.Stereotype.MODULE, self.module_config["view_finder"])
+        self.eps = float(self.config["eps"])
+        self.enable_global_scanned_feat = self.config["global_scanned_feat"]
+        
+    def forward(self, data):
+        mode = data["mode"]
+        
+        if mode == namespace.Mode.TRAIN:
+            return self.forward_train(data)
+        elif mode == namespace.Mode.TEST:
+            return self.forward_test(data)
+        else:
+            Log.error("Unknown mode: {}".format(mode), True)
+    
+    def pertube_data(self, gt_delta_9d):
+        bs = gt_delta_9d.shape[0]
+        random_t = torch.rand(bs, device=gt_delta_9d.device) * (1. - self.eps) + self.eps
+        random_t = random_t.unsqueeze(-1)
+        mu, std = self.view_finder.marginal_prob(gt_delta_9d, random_t)
+        std = std.view(-1, 1)
+        z = torch.randn_like(gt_delta_9d)
+        perturbed_x = mu + z * std
+        target_score = - z * std / (std ** 2)
+        return perturbed_x, random_t, target_score, std
+    
+    def forward_train(self, data):
+        main_feat = self.get_main_feat(data)
+        ''' get std '''
+        best_to_world_pose_9d_batch = data["best_to_world_pose_9d"]
+        perturbed_x, random_t, target_score, std = self.pertube_data(best_to_world_pose_9d_batch)
+        input_data = {
+            "sampled_pose": perturbed_x,
+            "t": random_t,
+            "main_feat": main_feat,
+        }
+        estimated_score = self.view_finder(input_data)
+        output = {
+            "estimated_score": estimated_score,
+            "target_score": target_score,
+            "std": std
+        }
+        return output
+    
+    def forward_test(self,data):
+        main_feat = self.get_main_feat(data)
+        estimated_delta_rot_9d, in_process_sample = self.view_finder.next_best_view(main_feat)
+        result = {
+            "pred_pose_9d": estimated_delta_rot_9d,
+            "in_process_sample": in_process_sample
+        }
+        return result
+    
+    
+    def get_main_feat(self, data):
+        scanned_pts_batch = data['scanned_pts']
+        scanned_n_to_world_pose_9d_batch = data['scanned_n_to_world_pose_9d']
+        
+
+        device = next(self.parameters()).device
+
+        
+
+        pts_feat_seq_list = []
+        pose_feat_seq_list = []
+        
+        for scanned_pts,scanned_n_to_world_pose_9d in zip(scanned_pts_batch,scanned_n_to_world_pose_9d_batch):
+            
+            scanned_pts = scanned_pts.to(device)
+            scanned_n_to_world_pose_9d = scanned_n_to_world_pose_9d.to(device)
+            pts_feat_seq_list.append(self.pts_encoder.encode_points(scanned_pts))
+            pose_feat_seq_list.append(self.pose_encoder.encode_pose(scanned_n_to_world_pose_9d))
+
+        main_feat = self.seq_encoder.encode_sequence(pts_feat_seq_list, pose_feat_seq_list)
+        
+        if self.enable_global_scanned_feat:
+            combined_scanned_pts_batch = data['combined_scanned_pts']   
+            global_scanned_feat = self.pts_encoder.encode_points(combined_scanned_pts_batch)
+            main_feat = torch.cat([main_feat, global_scanned_feat], dim=-1)
+        
+        
+        if torch.isnan(main_feat).any():
+            Log.error("nan in main_feat", True)
+        
+        return main_feat
+    

core/nbv_dataset.py

@@ -7,12 +7,13 @@ from PytorchBoot.utils.log_util import Log
 import torch
 import os
 import sys
-sys.path.append(r"/media/hofee/data/project/python/nbv_reconstruction/nbv_reconstruction")
+import time
+
+sys.path.append(r"/data/hofee/project/nbv_rec/nbv_reconstruction")
 
 from utils.data_load import DataLoadUtil
 from utils.pose import PoseUtil
 from utils.pts import PtsUtil
-from utils.reconstruction import ReconstructionUtil
 
 
 @stereotype.dataset("nbv_reconstruction_dataset")
@@ -30,9 +31,8 @@ class NBVReconstructionDataset(BaseDataset):
         self.cache = config.get("cache")
         self.load_from_preprocess = config.get("load_from_preprocess", False)
 
-
         if self.type == namespace.Mode.TEST:
-            self.model_dir = config["model_dir"]
+            #self.model_dir = config["model_dir"]
             self.filter_degree = config["filter_degree"]
         if self.type == namespace.Mode.TRAIN:
             scale_ratio = 1
@@ -41,9 +41,7 @@ class NBVReconstructionDataset(BaseDataset):
             expr_root = ConfigManager.get("runner", "experiment", "root_dir")
             expr_name = ConfigManager.get("runner", "experiment", "name")
             self.cache_dir = os.path.join(expr_root, expr_name, "cache")
-            #self.preprocess_cache()
-        
-        
+            # self.preprocess_cache()
 
     def load_scene_name_list(self):
         scene_name_list = []
@@ -52,24 +50,46 @@ class NBVReconstructionDataset(BaseDataset):
                 scene_name = line.strip()
                 scene_name_list.append(scene_name)
         return scene_name_list
-    
+
     def get_datalist(self):
         datalist = []
         for scene_name in self.scene_name_list:
-            label_path = DataLoadUtil.get_label_path_old(self.root_dir, scene_name)
-            label_data = DataLoadUtil.load_label(label_path)
-            for data_pair in label_data["data_pairs"]:
-                scanned_views = data_pair[0]
-                next_best_view = data_pair[1]
-                max_coverage_rate = label_data["max_coverage_rate"]
-                datalist.append(
-                    {
-                        "scanned_views": scanned_views,
-                        "next_best_view": next_best_view,
-                        "max_coverage_rate": max_coverage_rate,
-                        "scene_name": scene_name,
-                    }
+            seq_num = DataLoadUtil.get_label_num(self.root_dir, scene_name)
+            scene_max_coverage_rate = 0
+            max_coverage_rate_list = []
+
+            for seq_idx in range(seq_num):
+                label_path = DataLoadUtil.get_label_path(
+                    self.root_dir, scene_name, seq_idx
                 )
+                label_data = DataLoadUtil.load_label(label_path)
+                max_coverage_rate = label_data["max_coverage_rate"]
+                if max_coverage_rate > scene_max_coverage_rate:
+                    scene_max_coverage_rate = max_coverage_rate
+                max_coverage_rate_list.append(max_coverage_rate)
+                
+            if max_coverage_rate_list:
+                mean_coverage_rate = np.mean(max_coverage_rate_list)
+
+            for seq_idx in range(seq_num):
+                label_path = DataLoadUtil.get_label_path(
+                    self.root_dir, scene_name, seq_idx
+                )
+                label_data = DataLoadUtil.load_label(label_path)
+                if max_coverage_rate_list[seq_idx] > mean_coverage_rate - 0.1:
+                    for data_pair in label_data["data_pairs"]:
+                        scanned_views = data_pair[0]
+                        next_best_view = data_pair[1]
+                        datalist.append(
+                            {
+                                "scanned_views": scanned_views,
+                                "next_best_view": next_best_view,
+                                "seq_max_coverage_rate": max_coverage_rate,
+                                "scene_name": scene_name,
+                                "label_idx": seq_idx,
+                                "scene_max_coverage_rate": scene_max_coverage_rate,
+                            }
+                        )
         return datalist
 
     def preprocess_cache(self):
@@ -77,7 +97,7 @@ class NBVReconstructionDataset(BaseDataset):
         for item_idx in range(len(self.datalist)):
             self.__getitem__(item_idx)
         Log.success("finish preprocessing cache.")
-    
+
     def load_from_cache(self, scene_name, curr_frame_idx):
         cache_name = f"{scene_name}_{curr_frame_idx}.txt"
         cache_path = os.path.join(self.cache_dir, cache_name)
@@ -86,7 +106,7 @@ class NBVReconstructionDataset(BaseDataset):
             return data
         else:
             return None
-        
+
     def save_to_cache(self, scene_name, curr_frame_idx, data):
         cache_name = f"{scene_name}_{curr_frame_idx}.txt"
         cache_path = os.path.join(self.cache_dir, cache_name)
@@ -94,128 +114,153 @@ class NBVReconstructionDataset(BaseDataset):
             np.savetxt(cache_path, data)
         except Exception as e:
             Log.error(f"Save cache failed: {e}")
-            # ----- Debug Trace ----- #
-            import ipdb; ipdb.set_trace()
-            # ------------------------ #
-        
+
+    def voxel_downsample_with_mapping(self, point_cloud, voxel_size=0.003):
+        voxel_indices = np.floor(point_cloud / voxel_size).astype(np.int32)
+        unique_voxels, inverse, counts = np.unique(voxel_indices, axis=0, return_inverse=True, return_counts=True)
+        idx_sort = np.argsort(inverse)
+        idx_unique = idx_sort[np.cumsum(counts)-counts]
+        downsampled_points = point_cloud[idx_unique]
+        return downsampled_points, inverse
+
+
     def __getitem__(self, index):
         data_item_info = self.datalist[index]
         scanned_views = data_item_info["scanned_views"]
         nbv = data_item_info["next_best_view"]
-        max_coverage_rate = data_item_info["max_coverage_rate"]
+        max_coverage_rate = data_item_info["seq_max_coverage_rate"]
         scene_name = data_item_info["scene_name"]
-        scanned_views_pts, scanned_coverages_rate, scanned_n_to_world_pose = [], [], []
-      
+        (
+            scanned_views_pts,
+            scanned_coverages_rate,
+            scanned_n_to_world_pose,
+        ) = ([], [], [])
+        start_time = time.time()
+        start_indices = [0]
+        total_points = 0
         for view in scanned_views:
             frame_idx = view[0]
             coverage_rate = view[1]
             view_path = DataLoadUtil.get_path(self.root_dir, scene_name, frame_idx)
             cam_info = DataLoadUtil.load_cam_info(view_path, binocular=True)
-            n_to_world_pose = cam_info["cam_to_world"]
-            nR_to_world_pose = cam_info["cam_to_world_R"] 
-
-            if self.load_from_preprocess:
-                downsampled_target_point_cloud = DataLoadUtil.load_from_preprocessed_pts(view_path)
-            else:
-                cached_data = None
-                if self.cache:
-                    cached_data = self.load_from_cache(scene_name, frame_idx)
-                
-                if cached_data is None:
-                    print("load depth")
-                    depth_L, depth_R = DataLoadUtil.load_depth(view_path, cam_info['near_plane'], cam_info['far_plane'], binocular=True)
-                    point_cloud_L = DataLoadUtil.get_point_cloud(depth_L, cam_info['cam_intrinsic'], n_to_world_pose)['points_world']
-                    point_cloud_R = DataLoadUtil.get_point_cloud(depth_R, cam_info['cam_intrinsic'], nR_to_world_pose)['points_world']
-                
-                    point_cloud_L = PtsUtil.random_downsample_point_cloud(point_cloud_L, 65536)
-                    point_cloud_R = PtsUtil.random_downsample_point_cloud(point_cloud_R, 65536)
-                    overlap_points = DataLoadUtil.get_overlapping_points(point_cloud_L, point_cloud_R)
-                    downsampled_target_point_cloud = PtsUtil.random_downsample_point_cloud(overlap_points, self.pts_num)
-                    if self.cache:
-                        self.save_to_cache(scene_name, frame_idx, downsampled_target_point_cloud)
-                else:
-                    downsampled_target_point_cloud = cached_data
-                
+            
+            n_to_world_pose = cam_info["cam_to_world"]  
+            target_point_cloud = (
+                DataLoadUtil.load_from_preprocessed_pts(view_path)
+            )
+            downsampled_target_point_cloud = PtsUtil.random_downsample_point_cloud(
+                target_point_cloud, self.pts_num
+            )
             scanned_views_pts.append(downsampled_target_point_cloud)
-            scanned_coverages_rate.append(coverage_rate)          
-            n_to_world_6d = PoseUtil.matrix_to_rotation_6d_numpy(np.asarray(n_to_world_pose[:3,:3]))
-            n_to_world_trans = n_to_world_pose[:3,3]
+            scanned_coverages_rate.append(coverage_rate)
+            n_to_world_6d = PoseUtil.matrix_to_rotation_6d_numpy(
+                np.asarray(n_to_world_pose[:3, :3])
+            )
+            n_to_world_trans = n_to_world_pose[:3, 3]
             n_to_world_9d = np.concatenate([n_to_world_6d, n_to_world_trans], axis=0)
             scanned_n_to_world_pose.append(n_to_world_9d)
+            total_points += len(downsampled_target_point_cloud)
+            start_indices.append(total_points)
 
+
+        end_time = time.time()
+        #Log.info(f"load data time: {end_time - start_time}")
         nbv_idx, nbv_coverage_rate = nbv[0], nbv[1]
         nbv_path = DataLoadUtil.get_path(self.root_dir, scene_name, nbv_idx)
         cam_info = DataLoadUtil.load_cam_info(nbv_path)
         best_frame_to_world = cam_info["cam_to_world"]
+
+        best_to_world_6d = PoseUtil.matrix_to_rotation_6d_numpy(
+            np.asarray(best_frame_to_world[:3, :3])
+        )
+        best_to_world_trans = best_frame_to_world[:3, 3]
+        best_to_world_9d = np.concatenate(
+            [best_to_world_6d, best_to_world_trans], axis=0
+        )
         
-        best_to_world_6d = PoseUtil.matrix_to_rotation_6d_numpy(np.asarray(best_frame_to_world[:3,:3]))
-        best_to_world_trans = best_frame_to_world[:3,3]
-        best_to_world_9d = np.concatenate([best_to_world_6d, best_to_world_trans], axis=0)
+        combined_scanned_views_pts = np.concatenate(scanned_views_pts, axis=0)
+        voxel_downsampled_combined_scanned_pts_np, inverse = self.voxel_downsample_with_mapping(combined_scanned_views_pts, 0.003)
+        random_downsampled_combined_scanned_pts_np, random_downsample_idx = PtsUtil.random_downsample_point_cloud(voxel_downsampled_combined_scanned_pts_np, self.pts_num, require_idx=True)
+
+        all_idx_unique = np.arange(len(voxel_downsampled_combined_scanned_pts_np))
+        all_random_downsample_idx = all_idx_unique[random_downsample_idx]
+        scanned_pts_mask = []
+        for idx, start_idx in enumerate(start_indices):
+            if idx == len(start_indices) - 1:
+                break
+            end_idx = start_indices[idx+1]
+            view_inverse = inverse[start_idx:end_idx]
+            view_unique_downsampled_idx = np.unique(view_inverse)
+            view_unique_downsampled_idx_set = set(view_unique_downsampled_idx)
+            mask = np.array([idx in view_unique_downsampled_idx_set for idx in all_random_downsample_idx])
+            scanned_pts_mask.append(mask)
         data_item = {
-            "scanned_pts": np.asarray(scanned_views_pts,dtype=np.float32),
-            "scanned_coverage_rate": scanned_coverages_rate,
-            "scanned_n_to_world_pose_9d": np.asarray(scanned_n_to_world_pose,dtype=np.float32),
-            "best_coverage_rate": nbv_coverage_rate,
-            "best_to_world_pose_9d": np.asarray(best_to_world_9d,dtype=np.float32),
-            "max_coverage_rate": max_coverage_rate,
-            "scene_name": scene_name
+            "scanned_pts": np.asarray(scanned_views_pts, dtype=np.float32), # Ndarray(S x Nv x 3)
+            "combined_scanned_pts": np.asarray(random_downsampled_combined_scanned_pts_np, dtype=np.float32), # Ndarray(N x 3)
+            "scanned_pts_mask": np.asarray(scanned_pts_mask, dtype=np.bool), # Ndarray(N)
+            "scanned_coverage_rate": scanned_coverages_rate, # List(S): Float, range(0, 1)
+            "scanned_n_to_world_pose_9d": np.asarray(scanned_n_to_world_pose, dtype=np.float32), # Ndarray(S x 9)
+            "best_coverage_rate": nbv_coverage_rate, # Float, range(0, 1)
+            "best_to_world_pose_9d": np.asarray(best_to_world_9d, dtype=np.float32), # Ndarray(9)
+            "seq_max_coverage_rate": max_coverage_rate, # Float, range(0, 1)
+            "scene_name": scene_name, # String
         }
-        
-            
-        # if self.type == namespace.Mode.TEST:
-        #     diag = DataLoadUtil.get_bbox_diag(self.model_dir, scene_name)
-        #     voxel_threshold = diag*0.02
-        #     model_points_normals = DataLoadUtil.load_points_normals(self.root_dir, scene_name)
-        #     pts_list = []
-        #     for view in scanned_views:
-        #         frame_idx = view[0]
-        #         view_path = DataLoadUtil.get_path(self.root_dir, scene_name, frame_idx)
-        #         point_cloud = DataLoadUtil.get_target_point_cloud_world_from_path(view_path, binocular=True)
-        #         cam_params = DataLoadUtil.load_cam_info(view_path, binocular=True)
-        #         sampled_point_cloud = ReconstructionUtil.filter_points(point_cloud, model_points_normals, cam_pose=cam_params["cam_to_world"], voxel_size=voxel_threshold, theta=self.filter_degree)
-        #         pts_list.append(sampled_point_cloud)
-        #     nL_to_world_pose = cam_params["cam_to_world"]
-        #     nO_to_world_pose = cam_params["cam_to_world_O"]
-        #     nO_to_nL_pose = np.dot(np.linalg.inv(nL_to_world_pose), nO_to_world_pose)
-        #     data_item["scanned_target_pts_list"] = pts_list
-        #     data_item["model_points_normals"] = model_points_normals
-        #     data_item["voxel_threshold"] = voxel_threshold
-        #     data_item["filter_degree"] = self.filter_degree
-        #     data_item["scene_path"] = os.path.join(self.root_dir, scene_name)
-        #     data_item["first_frame_to_world"] = np.asarray(first_frame_to_world, dtype=np.float32)
-        #     data_item["nO_to_nL_pose"] = np.asarray(nO_to_nL_pose, dtype=np.float32)
+
         return data_item
 
     def __len__(self):
         return len(self.datalist)
-    
+
     def get_collate_fn(self):
         def collate_fn(batch):
             collate_data = {}
-            collate_data["scanned_pts"] = [torch.tensor(item['scanned_pts']) for item in batch]
-            collate_data["scanned_n_to_world_pose_9d"] = [torch.tensor(item['scanned_n_to_world_pose_9d']) for item in batch]
-            collate_data["best_to_world_pose_9d"] = torch.stack([torch.tensor(item['best_to_world_pose_9d']) for item in batch])
-            if "first_frame_to_world" in batch[0]:
-                collate_data["first_frame_to_world"] = torch.stack([torch.tensor(item["first_frame_to_world"]) for item in batch])
+            
+            ''' ------ Varialbe Length ------ '''
+            
+            collate_data["scanned_pts"] = [
+                torch.tensor(item["scanned_pts"]) for item in batch
+            ]
+            collate_data["scanned_n_to_world_pose_9d"] = [
+                torch.tensor(item["scanned_n_to_world_pose_9d"]) for item in batch
+            ]
+            collate_data["scanned_pts_mask"] = [
+                torch.tensor(item["scanned_pts_mask"]) for item in batch
+            ]
+            ''' ------ Fixed Length ------ '''
+            
+            collate_data["best_to_world_pose_9d"] = torch.stack(
+                [torch.tensor(item["best_to_world_pose_9d"]) for item in batch]
+            )
+            collate_data["combined_scanned_pts"] = torch.stack(
+                [torch.tensor(item["combined_scanned_pts"]) for item in batch]
+            )
+            
             for key in batch[0].keys():
-                if key not in ["scanned_pts", "scanned_n_to_world_pose_9d", "best_to_world_pose_9d", "first_frame_to_world"]:
+                if key not in [
+                    "scanned_pts",
+                    "scanned_n_to_world_pose_9d",
+                    "best_to_world_pose_9d",
+                    "combined_scanned_pts",
+                    "scanned_pts_mask",
+                ]:
                     collate_data[key] = [item[key] for item in batch]
             return collate_data
+
         return collate_fn
 
 
 # -------------- Debug ---------------- #
 if __name__ == "__main__":
     import torch
+
     seed = 0
     torch.manual_seed(seed)
     np.random.seed(seed)
     config = {
-        "root_dir": "/media/hofee/repository/nbv_reconstruction_data_512",
-        "model_dir": "/media/hofee/data/data/scaled_object_meshes",
+        "root_dir": "/data/hofee/nbv_rec_part2_preprocessed",
         "source": "nbv_reconstruction_dataset",
-        "split_file": "/media/hofee/data/project/python/nbv_reconstruction/sample_for_training/OmniObject3d_train.txt",
-        "load_from_preprocess": False,
+        "split_file": "/data/hofee/data/sample.txt",
+        "load_from_preprocess": True,
         "ratio": 0.5,
         "batch_size": 2,
         "filter_degree": 75,
@@ -225,41 +270,13 @@ if __name__ == "__main__":
     }
     ds = NBVReconstructionDataset(config)
     print(len(ds))
-    #ds.__getitem__(10)
+    # ds.__getitem__(10)
     dl = ds.get_loader(shuffle=True)
     for idx, data in enumerate(dl):
         data = ds.process_batch(data, "cuda:0")
         print(data)
         # ------  Debug Start ------
-        import ipdb;ipdb.set_trace()
+        import ipdb
+
+        ipdb.set_trace()
         # ------  Debug End ------
-    # 
-    # for idx, data in enumerate(dl):
-    #     cnt=0
-    #     print(data["scene_name"])
-    #     print(data["scanned_coverage_rate"])
-    #     print(data["best_coverage_rate"])
-    #     for pts in data["scanned_pts"][0]:
-    #         #np.savetxt(f"pts_{cnt}.txt", pts)
-    #         cnt+=1
-    #     #np.savetxt("best_pts.txt", best_pts)
-    #     for key, value in data.items():
-    #         if isinstance(value, torch.Tensor):
-    #             print(key, ":" ,value.shape)
-    #         else:
-    #             print(key, ":" ,len(value))
-    #         if key == "scanned_n_to_world_pose_9d":
-    #             for val in value:
-    #                 print(val.shape)
-    #         if key == "scanned_pts":
-    #             print("scanned_pts")
-    #             for val in value:
-    #                 print(val.shape)
-    #                 cnt = 0
-    #                 for v in val:
-    #                     import ipdb;ipdb.set_trace()
-    #                     np.savetxt(f"pts_{cnt}.txt", v)
-    #                     cnt+=1
-            
-                
-    #     print()

core/pipeline.py

@@ -1,84 +1,137 @@
 import torch
+import time
 from torch import nn
 import PytorchBoot.namespace as namespace
 import PytorchBoot.stereotype as stereotype
 from PytorchBoot.factory.component_factory import ComponentFactory
 from PytorchBoot.utils import Log
 
+
 @stereotype.pipeline("nbv_reconstruction_pipeline")
 class NBVReconstructionPipeline(nn.Module):
     def __init__(self, config):
         super(NBVReconstructionPipeline, self).__init__()
         self.config = config
-        self.pts_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, config["pts_encoder"])
-        self.pose_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, config["pose_encoder"])
-        self.seq_encoder = ComponentFactory.create(namespace.Stereotype.MODULE, config["seq_encoder"])
-        self.view_finder = ComponentFactory.create(namespace.Stereotype.MODULE, config["view_finder"])
-        self.eps = 1e-5
+        self.module_config = config["modules"]
         
+        self.pts_encoder = ComponentFactory.create(
+            namespace.Stereotype.MODULE, self.module_config["pts_encoder"]
+        )
+        self.pose_encoder = ComponentFactory.create(
+            namespace.Stereotype.MODULE, self.module_config["pose_encoder"]
+        )
+        self.seq_encoder = ComponentFactory.create(
+            namespace.Stereotype.MODULE, self.module_config["seq_encoder"]
+        )
+        self.view_finder = ComponentFactory.create(
+            namespace.Stereotype.MODULE, self.module_config["view_finder"]
+        )
+        
+
+        self.eps = float(self.config["eps"])
+
     def forward(self, data):
         mode = data["mode"]
-        
+
         if mode == namespace.Mode.TRAIN:
             return self.forward_train(data)
         elif mode == namespace.Mode.TEST:
             return self.forward_test(data)
         else:
             Log.error("Unknown mode: {}".format(mode), True)
-    
+
     def pertube_data(self, gt_delta_9d):
         bs = gt_delta_9d.shape[0]
-        random_t = torch.rand(bs, device=gt_delta_9d.device) * (1. - self.eps) + self.eps
+        random_t = (
+            torch.rand(bs, device=gt_delta_9d.device) * (1.0 - self.eps) + self.eps
+        )
         random_t = random_t.unsqueeze(-1)
         mu, std = self.view_finder.marginal_prob(gt_delta_9d, random_t)
         std = std.view(-1, 1)
         z = torch.randn_like(gt_delta_9d)
         perturbed_x = mu + z * std
-        target_score = - z * std / (std ** 2)
+        target_score = -z * std / (std**2)
         return perturbed_x, random_t, target_score, std
-    
+
     def forward_train(self, data):
-        seq_feat = self.get_seq_feat(data)
-        ''' get std '''
+        main_feat = self.get_main_feat(data)
+        """ get std """
         best_to_world_pose_9d_batch = data["best_to_world_pose_9d"]
-        perturbed_x, random_t, target_score, std = self.pertube_data(best_to_world_pose_9d_batch)
+        perturbed_x, random_t, target_score, std = self.pertube_data(
+            best_to_world_pose_9d_batch
+        )
         input_data = {
             "sampled_pose": perturbed_x,
             "t": random_t,
-            "seq_feat": seq_feat,
+            "main_feat": main_feat,
         }
         estimated_score = self.view_finder(input_data)
         output = {
             "estimated_score": estimated_score,
             "target_score": target_score,
-            "std": std
+            "std": std,
         }
         return output
-    
-    def forward_test(self,data):
-        seq_feat = self.get_seq_feat(data)
-        estimated_delta_rot_9d, in_process_sample = self.view_finder.next_best_view(seq_feat)
+
+    def forward_test(self, data):
+        main_feat = self.get_main_feat(data)
+        estimated_delta_rot_9d, in_process_sample = self.view_finder.next_best_view(
+            main_feat
+        )
         result = {
             "pred_pose_9d": estimated_delta_rot_9d,
-            "in_process_sample": in_process_sample
+            "in_process_sample": in_process_sample,
         }
         return result
-    
-    def get_seq_feat(self, data):
-        scanned_pts_batch = data['scanned_pts']
-        scanned_n_to_world_pose_9d_batch = data['scanned_n_to_world_pose_9d']
-        pts_feat_seq_list = []
-        pose_feat_seq_list = []
+
+    def get_main_feat(self, data):
+        scanned_n_to_world_pose_9d_batch = data[
+            "scanned_n_to_world_pose_9d"
+        ]  # List(B): Tensor(S x 9)
+
+        scanned_pts_mask_batch = data["scanned_pts_mask"] # List(B): Tensor(N)
+
         device = next(self.parameters()).device
-        for scanned_pts,scanned_n_to_world_pose_9d in zip(scanned_pts_batch,scanned_n_to_world_pose_9d_batch):
+
+        embedding_list_batch = []
+
+        combined_scanned_pts_batch = data["combined_scanned_pts"]  # Tensor(B x N x 3)
+        global_scanned_feat, per_point_feat_batch = self.pts_encoder.encode_points(
+            combined_scanned_pts_batch, require_per_point_feat=True
+        )  # global_scanned_feat: Tensor(B x Dg)
+        batch_size = len(scanned_n_to_world_pose_9d_batch)
+        for i in range(batch_size):
+            seq_len = len(scanned_n_to_world_pose_9d_batch[i])
+            scanned_n_to_world_pose_9d = scanned_n_to_world_pose_9d_batch[i].to(device)  # Tensor(S x 9)
+            scanned_pts_mask = scanned_pts_mask_batch[i] # Tensor(S x N)
+            per_point_feat = per_point_feat_batch[i] # Tensor(N x Dp)
+            partial_point_feat_seq = []
+            for j in range(seq_len):
+                partial_per_point_feat = per_point_feat[scanned_pts_mask[j]]
+                if partial_per_point_feat.shape[0] == 0:
+                    partial_point_feat = torch.zeros(per_point_feat.shape[1], device=device)
+                else:
+                    partial_point_feat = torch.mean(partial_per_point_feat, dim=0) # Tensor(Dp)
+                partial_point_feat_seq.append(partial_point_feat)
+            partial_point_feat_seq = torch.stack(partial_point_feat_seq, dim=0) # Tensor(S x Dp)
             
-            scanned_pts = scanned_pts.to(device)
-            scanned_n_to_world_pose_9d = scanned_n_to_world_pose_9d.to(device)
-            pts_feat_seq_list.append(self.pts_encoder.encode_points(scanned_pts))
-            pose_feat_seq_list.append(self.pose_encoder.encode_pose(scanned_n_to_world_pose_9d))
+            pose_feat_seq = self.pose_encoder.encode_pose(scanned_n_to_world_pose_9d)  # Tensor(S x Dp) 
+
+            seq_embedding = torch.cat([partial_point_feat_seq, pose_feat_seq], dim=-1)
+            
+            embedding_list_batch.append(seq_embedding) # List(B): Tensor(S x (Dp))
         
-        seq_feat = self.seq_encoder.encode_sequence(pts_feat_seq_list, pose_feat_seq_list)
-        if torch.isnan(seq_feat).any():
-            Log.error("nan in seq_feat", True)
-        return seq_feat
-    
+        seq_feat = self.seq_encoder.encode_sequence(embedding_list_batch) # Tensor(B x Ds)
+        main_feat = torch.cat([seq_feat, global_scanned_feat], dim=-1) # Tensor(B x (Ds+Dg))
+
+        if torch.isnan(main_feat).any():
+            for i in range(len(main_feat)):
+                if torch.isnan(main_feat[i]).any():
+                    scanned_pts_mask = scanned_pts_mask_batch[i]
+                    Log.info(f"scanned_pts_mask shape: {scanned_pts_mask.shape}")
+                    Log.info(f"scanned_pts_mask sum: {scanned_pts_mask.sum()}")
+                    import ipdb
+                    ipdb.set_trace()
+            Log.error("nan in main_feat", True)
+
+        return main_feat

core/seq_dataset.py

@@ -6,7 +6,7 @@ from PytorchBoot.utils.log_util import Log
 import torch
 import os
 import sys
-sys.path.append(r"/media/hofee/data/project/python/nbv_reconstruction/nbv_reconstruction")
+sys.path.append(r"/home/data/hofee/project/nbv_rec/nbv_reconstruction")
 
 from utils.data_load import DataLoadUtil
 from utils.pose import PoseUtil
@@ -39,42 +39,32 @@ class SeqNBVReconstructionDataset(BaseDataset):
                 scene_name_list.append(scene_name)
         return scene_name_list
     
-    def get_datalist_new(self):
-        datalist = []
-        for scene_name in self.scene_name_list:
-            label_num = DataLoadUtil.get_label_num(self.root_dir, scene_name)
-            for i in range(label_num):
-                label_path = DataLoadUtil.get_label_path(self.root_dir, scene_name, i)
-                label_data = DataLoadUtil.load_label(label_path)
-                best_seq = label_data["best_sequence"]
-                max_coverage_rate = label_data["max_coverage_rate"]
-                first_frame = best_seq[0]
-                best_seq_len = len(best_seq)
-                datalist.append({
-                    "scene_name": scene_name,
-                    "first_frame": first_frame,
-                    "max_coverage_rate": max_coverage_rate,
-                    "best_seq_len": best_seq_len,
-                    "label_idx": i,
-                })
-        return datalist
-    
     def get_datalist(self):
         datalist = []
         for scene_name in self.scene_name_list:
-            label_path = DataLoadUtil.get_label_path_old(self.root_dir, scene_name)
+            seq_num = DataLoadUtil.get_label_num(self.root_dir, scene_name)
+            scene_max_coverage_rate = 0
+            scene_max_cr_idx = 0
+
+            for seq_idx in range(seq_num):
+                label_path = DataLoadUtil.get_label_path(self.root_dir, scene_name, seq_idx)
+                label_data = DataLoadUtil.load_label(label_path)
+                max_coverage_rate = label_data["max_coverage_rate"]
+                if max_coverage_rate > scene_max_coverage_rate:
+                    scene_max_coverage_rate = max_coverage_rate
+                    scene_max_cr_idx = seq_idx
+        
+            label_path = DataLoadUtil.get_label_path(self.root_dir, scene_name, scene_max_cr_idx)
             label_data = DataLoadUtil.load_label(label_path)
-            best_seq = label_data["best_sequence"]
-            max_coverage_rate = label_data["max_coverage_rate"]
-            first_frame = best_seq[0]
-            best_seq_len = len(best_seq)
+            first_frame = label_data["best_sequence"][0]
+            best_seq_len = len(label_data["best_sequence"])
             datalist.append({
-                "scene_name": scene_name,
-                "first_frame": first_frame,
-                "max_coverage_rate": max_coverage_rate,
-                "best_seq_len": best_seq_len,
-                "best_seq": best_seq,
-            })
+                    "scene_name": scene_name,
+                    "first_frame": first_frame,
+                    "max_coverage_rate": scene_max_coverage_rate,
+                    "best_seq_len": best_seq_len,
+                    "label_idx": scene_max_cr_idx,
+                })
         return datalist
 
     def __getitem__(self, index):
@@ -84,24 +74,12 @@ class SeqNBVReconstructionDataset(BaseDataset):
         max_coverage_rate = data_item_info["max_coverage_rate"]
         scene_name = data_item_info["scene_name"]
         first_cam_info = DataLoadUtil.load_cam_info(DataLoadUtil.get_path(self.root_dir, scene_name, first_frame_idx), binocular=True)
-       
         first_view_path = DataLoadUtil.get_path(self.root_dir, scene_name, first_frame_idx)
         first_left_cam_pose = first_cam_info["cam_to_world"]
-        first_right_cam_pose = first_cam_info["cam_to_world_R"]
         first_center_cam_pose = first_cam_info["cam_to_world_O"]
-        if self.load_from_preprocess:
-            first_downsampled_target_point_cloud = DataLoadUtil.load_from_preprocessed_pts(first_view_path)
-        else:    
-            first_depth_L, first_depth_R = DataLoadUtil.load_depth(first_view_path, first_cam_info['near_plane'], first_cam_info['far_plane'], binocular=True)
-            
-            first_point_cloud_L = DataLoadUtil.get_point_cloud(first_depth_L, first_cam_info['cam_intrinsic'], first_left_cam_pose)['points_world']
-            first_point_cloud_R = DataLoadUtil.get_point_cloud(first_depth_R, first_cam_info['cam_intrinsic'], first_right_cam_pose)['points_world']
-            
-            first_point_cloud_L = PtsUtil.random_downsample_point_cloud(first_point_cloud_L, 65536)
-            first_point_cloud_R = PtsUtil.random_downsample_point_cloud(first_point_cloud_R, 65536)
-            first_overlap_points = DataLoadUtil.get_overlapping_points(first_point_cloud_L, first_point_cloud_R)
-            first_downsampled_target_point_cloud = PtsUtil.random_downsample_point_cloud(first_overlap_points, self.pts_num)
-     
+        first_target_point_cloud = DataLoadUtil.load_from_preprocessed_pts(first_view_path)
+        first_pts_num = first_target_point_cloud.shape[0]
+        first_downsampled_target_point_cloud = PtsUtil.random_downsample_point_cloud(first_target_point_cloud, self.pts_num)
         first_to_world_rot_6d = PoseUtil.matrix_to_rotation_6d_numpy(np.asarray(first_left_cam_pose[:3,:3]))
         first_to_world_trans = first_left_cam_pose[:3,3]
         first_to_world_9d = np.concatenate([first_to_world_rot_6d, first_to_world_trans], axis=0)
@@ -110,8 +88,13 @@ class SeqNBVReconstructionDataset(BaseDataset):
         first_O_to_first_L_pose = np.dot(np.linalg.inv(first_left_cam_pose), first_center_cam_pose)
         scene_path = os.path.join(self.root_dir, scene_name)
         model_points_normals = DataLoadUtil.load_points_normals(self.root_dir, scene_name)
+
         data_item = {
+            "first_pts_num": np.asarray(
+                first_pts_num, dtype=np.int32
+            ),
             "first_pts": np.asarray([first_downsampled_target_point_cloud],dtype=np.float32),
+            "combined_scanned_pts": np.asarray(first_downsampled_target_point_cloud,dtype=np.float32),
             "first_to_world_9d": np.asarray([first_to_world_9d],dtype=np.float32),
             "scene_name": scene_name,
             "max_coverage_rate": max_coverage_rate,
@@ -134,8 +117,9 @@ class SeqNBVReconstructionDataset(BaseDataset):
             collate_data = {}
             collate_data["first_pts"] = [torch.tensor(item['first_pts']) for item in batch]
             collate_data["first_to_world_9d"] = [torch.tensor(item['first_to_world_9d']) for item in batch]
+            collate_data["combined_scanned_pts"] = torch.stack([torch.tensor(item['combined_scanned_pts']) for item in batch])
             for key in batch[0].keys():
-                if key not in ["first_pts", "first_to_world_9d"]:
+                if key not in ["first_pts", "first_to_world_9d", "combined_scanned_pts"]:
                     collate_data[key] = [item[key] for item in batch]
             return collate_data
         return collate_fn
@@ -147,15 +131,16 @@ if __name__ == "__main__":
     torch.manual_seed(seed)
     np.random.seed(seed)
     config = {
-        "root_dir": "/media/hofee/data/project/python/nbv_reconstruction/sample_for_training/scenes",
-        "split_file": "/media/hofee/data/project/python/nbv_reconstruction/sample_for_training/OmniObject3d_train.txt",
-        "model_dir": "/media/hofee/data/data/scaled_object_meshes",
-        "ratio": 0.5,
+        "root_dir": "/home/data/hofee/project/nbv_rec/data/nbv_rec_data_512_preproc_npy",
+        "split_file": "/home/data/hofee/project/nbv_rec/data/OmniObject3d_train.txt",
+        "model_dir": "/home/data/hofee/project/nbv_rec/data/scaled_object_meshes",
+        "ratio": 0.005,
         "batch_size": 2,
         "filter_degree": 75,
         "num_workers": 0,
         "pts_num": 32684,
         "type": namespace.Mode.TEST,
+        "load_from_preprocess": True
     }
     ds = SeqNBVReconstructionDataset(config)
     print(len(ds))

modules/func_lib/samplers.py

@@ -32,7 +32,7 @@ def cond_ode_sampler(
     init_x=None,
 ):
     pose_dim = PoseUtil.get_pose_dim(pose_mode)
-    batch_size = data["seq_feat"].shape[0]
+    batch_size = data["main_feat"].shape[0]
     init_x = (
         prior((batch_size, pose_dim), T=T).to(device)
         if init_x is None

modules/gf_view_finder.py

@@ -80,13 +80,13 @@ class GradientFieldViewFinder(nn.Module):
         """
         Args:
             data, dict {
-                'seq_feat': [bs, c]
+                'main_feat': [bs, c]
                 'pose_sample': [bs, pose_dim]
                 't': [bs, 1]
             }
         """
 
-        seq_feat = data['seq_feat']
+        main_feat = data['main_feat']
         sampled_pose = data['sampled_pose']
         t = data['t']
         t_feat = self.t_encoder(t.squeeze(1))
@@ -95,7 +95,7 @@ class GradientFieldViewFinder(nn.Module):
         if self.per_point_feature:
             raise NotImplementedError
         else:
-            total_feat = torch.cat([seq_feat, t_feat, pose_feat], dim=-1)
+            total_feat = torch.cat([main_feat, t_feat, pose_feat], dim=-1)
         _, std = self.marginal_prob_fn(total_feat, t)
 
         if self.regression_head == 'Rx_Ry_and_T':
@@ -134,9 +134,9 @@ class GradientFieldViewFinder(nn.Module):
 
         return in_process_sample, res
 
-    def next_best_view(self, seq_feat):
+    def next_best_view(self, main_feat):
         data = {
-            'seq_feat': seq_feat,
+            'main_feat': main_feat,
         }
         in_process_sample, res = self.sample(data)
         return res.to(dtype=torch.float32), in_process_sample

modules/pointnet_encoder.py

@@ -22,12 +22,10 @@ class PointNetEncoder(nn.Module):
         self.conv2 = torch.nn.Conv1d(64, 128, 1)
         self.conv3 = torch.nn.Conv1d(128, 512, 1)
         self.conv4 = torch.nn.Conv1d(512, self.out_dim , 1)
-        self.global_feat = config["global_feat"]
         if self.feature_transform:
             self.f_stn = STNkd(k=64)
 
     def forward(self, x):
-        n_pts = x.shape[2]
         trans = self.stn(x)
         x = x.transpose(2, 1)
         x = torch.bmm(x, trans)
@@ -46,20 +44,15 @@ class PointNetEncoder(nn.Module):
         x = self.conv4(x)
         x = torch.max(x, 2, keepdim=True)[0]
         x = x.view(-1, self.out_dim)
-        if self.global_feat:
-            return x
-        else:
-            x = x.view(-1, self.out_dim, 1).repeat(1, 1, n_pts)
-            return torch.cat([x, point_feat], 1)
+        return x, point_feat
 
-    def encode_points(self, pts):
+    def encode_points(self, pts, require_per_point_feat=False):
         pts = pts.transpose(2, 1)
-
-        if not self.global_feat:
-            pts_feature = self(pts).transpose(2, 1)
+        global_pts_feature, per_point_feature = self(pts)
+        if require_per_point_feat:
+            return global_pts_feature, per_point_feature.transpose(2, 1)
         else:
-            pts_feature = self(pts)
-        return pts_feature
+            return global_pts_feature
 
 class STNkd(nn.Module):
     def __init__(self, k=64):
@@ -102,21 +95,13 @@ if __name__ == "__main__":
     config = {
         "in_dim": 3,
         "out_dim": 1024,
-        "global_feat": True,
         "feature_transform": False
     }
-    pointnet_global = PointNetEncoder(config)
-    out = pointnet_global.encode_points(sim_data)
+    pointnet = PointNetEncoder(config)
+    out = pointnet.encode_points(sim_data)
 
     print("global feat", out.size())
 
-    config = {
-        "in_dim": 3,
-        "out_dim": 1024,
-        "global_feat": False,
-        "feature_transform": False
-    }
-
-    pointnet = PointNetEncoder(config)
-    out = pointnet.encode_points(sim_data)
+    out, per_point_out = pointnet.encode_points(sim_data, require_per_point_feat=True)
     print("point feat", out.size())
+    print("per point feat", per_point_out.size())

modules/pts_num_encoder.py

@@ -0,0 +1,20 @@
+from torch import nn
+import PytorchBoot.stereotype as stereotype
+
+@stereotype.module("pts_num_encoder")
+class PointsNumEncoder(nn.Module):
+    def __init__(self, config):
+        super(PointsNumEncoder, self).__init__()
+        self.config = config
+        out_dim = config["out_dim"]
+        self.act = nn.ReLU(True)
+
+        self.pts_num_encoder = nn.Sequential(
+            nn.Linear(1, out_dim),
+            self.act,
+            nn.Linear(out_dim, out_dim),
+            self.act,
+        )
+
+    def encode_pts_num(self, num_seq):
+        return self.pts_num_encoder(num_seq)

modules/transformer_seq_encoder.py

@@ -9,7 +9,7 @@ class TransformerSequenceEncoder(nn.Module):
     def __init__(self, config):
         super(TransformerSequenceEncoder, self).__init__()
         self.config = config
-        embed_dim = config["pts_embed_dim"] + config["pose_embed_dim"]
+        embed_dim = config["embed_dim"]
         encoder_layer = nn.TransformerEncoderLayer(
             d_model=embed_dim,
             nhead=config["num_heads"],
@@ -21,24 +21,19 @@ class TransformerSequenceEncoder(nn.Module):
         )
         self.fc = nn.Linear(embed_dim, config["output_dim"])
 
-    def encode_sequence(self, pts_embedding_list_batch, pose_embedding_list_batch):
-        combined_features_batch = []
+    def encode_sequence(self, embedding_list_batch):
+        
         lengths = []
+        
+        for embedding_list in embedding_list_batch:
+            lengths.append(len(embedding_list))
 
-        for pts_embedding_list, pose_embedding_list in zip(pts_embedding_list_batch, pose_embedding_list_batch):
-            combined_features = [
-                torch.cat((pts_embed, pose_embed), dim=-1)
-                for pts_embed, pose_embed in zip(pts_embedding_list, pose_embedding_list)
-            ]
-            combined_features_batch.append(torch.stack(combined_features))
-            lengths.append(len(combined_features))
-
-        combined_tensor = pad_sequence(combined_features_batch, batch_first=True)  # Shape: [batch_size, max_seq_len, embed_dim]
+        embedding_tensor = pad_sequence(embedding_list_batch, batch_first=True)  # Shape: [batch_size, max_seq_len, embed_dim]
 
         max_len = max(lengths)
-        padding_mask = torch.tensor([([0] * length + [1] * (max_len - length)) for length in lengths], dtype=torch.bool).to(combined_tensor.device)
+        padding_mask = torch.tensor([([0] * length + [1] * (max_len - length)) for length in lengths], dtype=torch.bool).to(embedding_tensor.device)
 
-        transformer_output = self.transformer_encoder(combined_tensor, src_key_padding_mask=padding_mask)
+        transformer_output = self.transformer_encoder(embedding_tensor, src_key_padding_mask=padding_mask)
         final_feature = transformer_output.mean(dim=1)
         final_output = self.fc(final_feature)
 
@@ -47,26 +42,22 @@ class TransformerSequenceEncoder(nn.Module):
 
 if __name__ == "__main__":
     config = {
-        "pts_embed_dim": 1024, 
-        "pose_embed_dim": 256,
+        "embed_dim": 256,
         "num_heads": 4,  
         "ffn_dim": 256,
         "num_layers": 3, 
-        "output_dim": 2048, 
+        "output_dim": 1024, 
     }
 
     encoder = TransformerSequenceEncoder(config)
     seq_len = [5, 8, 9, 4]
     batch_size = 4
 
-    pts_embedding_list_batch = [
-        torch.randn(seq_len[idx], config["pts_embed_dim"]) for idx in range(batch_size)
-    ]
-    pose_embedding_list_batch = [
-        torch.randn(seq_len[idx], config["pose_embed_dim"]) for idx in range(batch_size)
+    embedding_list_batch = [
+        torch.randn(seq_len[idx], config["embed_dim"]) for idx in range(batch_size)
     ]
     output_feature = encoder.encode_sequence(
-        pts_embedding_list_batch, pose_embedding_list_batch
+        embedding_list_batch
     )
     print("Encoded Feature:", output_feature)
     print("Feature Shape:", output_feature.shape)

preprocess/clean_preprocessed_data.py

@@ -0,0 +1,43 @@
+import os
+import shutil
+
+def clean_scene_data(root, scene):
+    # 清理目标点云数据
+    pts_dir = os.path.join(root, scene, "pts")
+    if os.path.exists(pts_dir):
+        shutil.rmtree(pts_dir)
+        print(f"已删除 {pts_dir}")
+
+    # 清理法线数据
+    nrm_dir = os.path.join(root, scene, "nrm")
+    if os.path.exists(nrm_dir):
+        shutil.rmtree(nrm_dir)
+        print(f"已删除 {nrm_dir}")
+
+    # 清理扫描点索引数据
+    scan_points_indices_dir = os.path.join(root, scene, "scan_points_indices")
+    if os.path.exists(scan_points_indices_dir):
+        shutil.rmtree(scan_points_indices_dir)
+        print(f"已删除 {scan_points_indices_dir}")
+
+    # 删除扫描点数据文件
+    scan_points_file = os.path.join(root, scene, "scan_points.txt")
+    if os.path.exists(scan_points_file):
+        os.remove(scan_points_file)
+        print(f"已删除 {scan_points_file}")
+
+def clean_all_scenes(root, scene_list):
+    for idx, scene in enumerate(scene_list):
+        print(f"正在清理场景 {scene} ({idx+1}/{len(scene_list)})")
+        clean_scene_data(root, scene)
+
+if __name__ == "__main__":
+    root = r"c:\Document\Local Project\nbv_rec\nbv_reconstruction\temp"
+    scene_list = os.listdir(root)
+    from_idx = 0
+    to_idx = len(scene_list)
+    print(f"正在清理场景 {scene_list[from_idx:to_idx]}")
+
+    clean_all_scenes(root, scene_list[from_idx:to_idx])
+    print("清理完成")
+

preprocess/pack_preprocessed_data.py

@@ -0,0 +1,48 @@
+import os
+import shutil
+
+def pack_scene_data(root, scene, output_dir):
+    scene_dir = os.path.join(output_dir, scene)
+    if not os.path.exists(scene_dir):
+        os.makedirs(scene_dir)
+
+    pts_dir = os.path.join(root, scene, "pts")
+    if os.path.exists(pts_dir):
+        shutil.move(pts_dir, os.path.join(scene_dir, "pts"))
+        
+    scan_points_indices_dir = os.path.join(root, scene, "scan_points_indices")
+    if os.path.exists(scan_points_indices_dir):
+        shutil.move(scan_points_indices_dir, os.path.join(scene_dir, "scan_points_indices")) 
+
+    scan_points_file = os.path.join(root, scene, "scan_points.txt")
+    if os.path.exists(scan_points_file):
+        shutil.move(scan_points_file, os.path.join(scene_dir, "scan_points.txt"))
+
+    model_pts_nrm_file = os.path.join(root, scene, "points_and_normals.txt")
+    if os.path.exists(model_pts_nrm_file):
+        shutil.move(model_pts_nrm_file, os.path.join(scene_dir, "points_and_normals.txt"))
+
+    camera_dir = os.path.join(root, scene, "camera_params")
+    if os.path.exists(camera_dir):
+        shutil.move(camera_dir, os.path.join(scene_dir, "camera_params"))
+
+    scene_info_file = os.path.join(root, scene, "scene_info.json")
+    if os.path.exists(scene_info_file):
+        shutil.move(scene_info_file, os.path.join(scene_dir, "scene_info.json"))
+
+def pack_all_scenes(root, scene_list, output_dir):
+    for idx, scene in enumerate(scene_list):
+        print(f"正在打包场景 {scene} ({idx+1}/{len(scene_list)})")
+        pack_scene_data(root, scene, output_dir)
+
+if __name__ == "__main__":
+    root = r"H:\AI\Datasets\nbv_rec_part2"
+    output_dir = r"H:\AI\Datasets\scene_info_part2"
+    scene_list = os.listdir(root)
+    from_idx = 0
+    to_idx = len(scene_list)
+    print(f"正在打包场景 {scene_list[from_idx:to_idx]}")
+
+    pack_all_scenes(root, scene_list[from_idx:to_idx], output_dir)
+    print("打包完成")
+

preprocess/pack_upload_data.py

@@ -0,0 +1,41 @@
+import os
+import shutil
+
+def pack_scene_data(root, scene, output_dir):
+    scene_dir = os.path.join(output_dir, scene)
+    if not os.path.exists(scene_dir):
+        os.makedirs(scene_dir)
+
+    pts_dir = os.path.join(root, scene, "pts")
+    if os.path.exists(pts_dir):
+        shutil.move(pts_dir, os.path.join(scene_dir, "pts"))
+
+    camera_dir = os.path.join(root, scene, "camera_params")
+    if os.path.exists(camera_dir):
+        shutil.move(camera_dir, os.path.join(scene_dir, "camera_params"))
+
+    scene_info_file = os.path.join(root, scene, "scene_info.json")
+    if os.path.exists(scene_info_file):
+        shutil.move(scene_info_file, os.path.join(scene_dir, "scene_info.json"))
+
+    label_dir = os.path.join(root, scene, "label")
+    if os.path.exists(label_dir):
+        shutil.move(label_dir, os.path.join(scene_dir, "label"))
+
+
+def pack_all_scenes(root, scene_list, output_dir):
+    for idx, scene in enumerate(scene_list):
+        print(f"packing {scene} ({idx+1}/{len(scene_list)})")
+        pack_scene_data(root, scene, output_dir)
+
+if __name__ == "__main__":
+    root = r"H:\AI\Datasets\nbv_rec_part2"
+    output_dir = r"H:\AI\Datasets\upload_part2"
+    scene_list = os.listdir(root)
+    from_idx = 0
+    to_idx = len(scene_list)
+    print(f"packing {scene_list[from_idx:to_idx]}")
+
+    pack_all_scenes(root, scene_list[from_idx:to_idx], output_dir)
+    print("packing done")
+

preprocess/preprocessor.py

@@ -0,0 +1,185 @@
+import os
+import numpy as np
+import time
+import sys
+np.random.seed(0)
+sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from utils.reconstruction import ReconstructionUtil
+from utils.data_load import DataLoadUtil
+from utils.pts import PtsUtil
+
+def save_np_pts(path, pts: np.ndarray, file_type="txt"):
+    if file_type == "txt":
+        np.savetxt(path, pts)
+    else:
+        np.save(path, pts)
+
+
+def save_target_points(root, scene, frame_idx, target_points: np.ndarray, file_type="txt"):
+    pts_path = os.path.join(root,scene, "pts", f"{frame_idx}.{file_type}")
+    if not os.path.exists(os.path.join(root,scene, "pts")):
+        os.makedirs(os.path.join(root,scene, "pts"))
+    save_np_pts(pts_path, target_points, file_type)
+
+def save_target_normals(root, scene, frame_idx, target_normals: np.ndarray, file_type="txt"):
+    pts_path = os.path.join(root,scene, "nrm", f"{frame_idx}.{file_type}")
+    if not os.path.exists(os.path.join(root,scene, "nrm")):
+        os.makedirs(os.path.join(root,scene, "nrm"))
+    save_np_pts(pts_path, target_normals, file_type)
+    
+def save_scan_points_indices(root, scene, frame_idx, scan_points_indices: np.ndarray, file_type="txt"):
+    indices_path = os.path.join(root,scene, "scan_points_indices", f"{frame_idx}.{file_type}")
+    if not os.path.exists(os.path.join(root,scene, "scan_points_indices")):
+        os.makedirs(os.path.join(root,scene, "scan_points_indices"))
+    save_np_pts(indices_path, scan_points_indices, file_type)
+
+def save_scan_points(root, scene, scan_points: np.ndarray):
+    scan_points_path = os.path.join(root,scene, "scan_points.txt")
+    save_np_pts(scan_points_path, scan_points)
+
+def get_world_points(depth, mask, cam_intrinsic, cam_extrinsic, random_downsample_N):
+    z = depth[mask]
+    i, j = np.nonzero(mask)
+    x = (j - cam_intrinsic[0, 2]) * z / cam_intrinsic[0, 0]
+    y = (i - cam_intrinsic[1, 2]) * z / cam_intrinsic[1, 1]
+
+    points_camera = np.stack((x, y, z), axis=-1).reshape(-1, 3)
+    sampled_target_points = PtsUtil.random_downsample_point_cloud(
+                points_camera, random_downsample_N
+            )
+    points_camera_aug = np.concatenate((sampled_target_points, np.ones((sampled_target_points.shape[0], 1))), axis=-1)
+    points_camera_world = np.dot(cam_extrinsic, points_camera_aug.T).T[:, :3]
+    
+    return points_camera_world
+
+def get_world_points_and_normal(depth, mask, normal, cam_intrinsic, cam_extrinsic, random_downsample_N):
+    z = depth[mask]
+    i, j = np.nonzero(mask)
+    x = (j - cam_intrinsic[0, 2]) * z / cam_intrinsic[0, 0]
+    y = (i - cam_intrinsic[1, 2]) * z / cam_intrinsic[1, 1]
+
+    points_camera = np.stack((x, y, z), axis=-1).reshape(-1, 3)
+    normal_camera = normal[mask].reshape(-1, 3)
+    sampled_target_points, idx = PtsUtil.random_downsample_point_cloud(
+                points_camera, random_downsample_N, require_idx=True
+            )
+    if len(sampled_target_points) == 0:
+        return np.zeros((0, 3)), np.zeros((0, 3))
+    sampled_normal_camera  = normal_camera[idx]
+
+    points_camera_aug = np.concatenate((sampled_target_points, np.ones((sampled_target_points.shape[0], 1))), axis=-1)
+    points_camera_world = np.dot(cam_extrinsic, points_camera_aug.T).T[:, :3]
+    
+    return points_camera_world, sampled_normal_camera
+
+def get_scan_points_indices(scan_points, mask, display_table_mask_label, cam_intrinsic, cam_extrinsic):
+    scan_points_homogeneous = np.hstack((scan_points, np.ones((scan_points.shape[0], 1))))
+    points_camera = np.dot(np.linalg.inv(cam_extrinsic), scan_points_homogeneous.T).T[:, :3]
+    points_image_homogeneous = np.dot(cam_intrinsic, points_camera.T).T
+    points_image_homogeneous /= points_image_homogeneous[:, 2:]
+    pixel_x = points_image_homogeneous[:, 0].astype(int)
+    pixel_y = points_image_homogeneous[:, 1].astype(int)
+    h, w = mask.shape[:2]
+    valid_indices = (pixel_x >= 0) & (pixel_x < w) & (pixel_y >= 0) & (pixel_y < h)
+    mask_colors = mask[pixel_y[valid_indices], pixel_x[valid_indices]]
+    selected_points_indices = np.where((mask_colors == display_table_mask_label).all(axis=-1))[0]
+    selected_points_indices = np.where(valid_indices)[0][selected_points_indices]
+    return selected_points_indices
+    
+
+def save_scene_data(root, scene, scene_idx=0, scene_total=1,file_type="txt"):
+    
+    ''' configuration '''
+    target_mask_label = (0, 255, 0)
+    display_table_mask_label=(0, 0, 255)
+    random_downsample_N = 32768
+    voxel_size=0.003
+    filter_degree = 75
+    min_z = 0.2
+    max_z = 0.5
+    
+    ''' scan points '''
+    display_table_info = DataLoadUtil.get_display_table_info(root, scene)
+    radius = display_table_info["radius"]
+    
+    scan_points = np.asarray(ReconstructionUtil.generate_scan_points(display_table_top=0,display_table_radius=radius))
+    
+    ''' read frame data(depth|mask|normal) '''
+    frame_num = DataLoadUtil.get_scene_seq_length(root, scene)
+    for frame_id in range(frame_num):
+        print(f"[scene({scene_idx}/{scene_total})|frame({frame_id}/{frame_num})]Processing {scene} frame {frame_id}")
+        path = DataLoadUtil.get_path(root, scene, frame_id)
+        cam_info = DataLoadUtil.load_cam_info(path, binocular=True)
+        depth_L, depth_R = DataLoadUtil.load_depth(
+                path, cam_info["near_plane"], 
+                cam_info["far_plane"], 
+                binocular=True
+            )
+        mask_L, mask_R = DataLoadUtil.load_seg(path, binocular=True)
+        normal_L = DataLoadUtil.load_normal(path, binocular=True, left_only=True)
+        ''' target points '''
+        mask_img_L = mask_L
+        mask_img_R = mask_R
+
+        target_mask_img_L = (mask_L == target_mask_label).all(axis=-1)
+        target_mask_img_R = (mask_R == target_mask_label).all(axis=-1)
+
+        
+        sampled_target_points_L, sampled_target_normal_L = get_world_points_and_normal(depth_L,target_mask_img_L,normal_L, cam_info["cam_intrinsic"], cam_info["cam_to_world"], random_downsample_N)
+        sampled_target_points_R = get_world_points(depth_R, target_mask_img_R, cam_info["cam_intrinsic"], cam_info["cam_to_world_R"], random_downsample_N)
+
+
+        has_points = sampled_target_points_L.shape[0] > 0 and sampled_target_points_R.shape[0] > 0
+        if has_points:
+            target_points, overlap_idx = PtsUtil.get_overlapping_points(
+                    sampled_target_points_L, sampled_target_points_R, voxel_size, require_idx=True
+                )
+            sampled_target_normal_L = sampled_target_normal_L[overlap_idx]
+
+        if has_points:
+            has_points = target_points.shape[0] > 0
+
+        if has_points:
+            target_points, target_normals = PtsUtil.filter_points(
+                target_points, sampled_target_normal_L, cam_info["cam_to_world"], theta_limit = filter_degree, z_range=(min_z, max_z)
+                )
+
+
+        ''' scan points indices '''
+        scan_points_indices_L = get_scan_points_indices(scan_points, mask_img_L, display_table_mask_label, cam_info["cam_intrinsic"], cam_info["cam_to_world"]) 
+        scan_points_indices_R = get_scan_points_indices(scan_points, mask_img_R, display_table_mask_label, cam_info["cam_intrinsic"], cam_info["cam_to_world_R"])
+        scan_points_indices = np.intersect1d(scan_points_indices_L, scan_points_indices_R)
+        
+        if not has_points:
+            target_points = np.zeros((0, 3))
+            target_normals = np.zeros((0, 3))
+            
+        save_target_points(root, scene, frame_id, target_points, file_type=file_type)
+        save_target_normals(root, scene, frame_id, target_normals, file_type=file_type)
+        save_scan_points_indices(root, scene, frame_id, scan_points_indices, file_type=file_type)
+    
+    save_scan_points(root, scene, scan_points) # The "done" flag of scene preprocess
+
+
+if __name__ == "__main__":
+    #root = "/media/hofee/repository/new_data_with_normal"
+    root = r"H:\AI\Datasets\nbv_rec_part2"
+    scene_list = os.listdir(root)
+    from_idx = 0 # 1000
+    to_idx = 600 # 1500
+
+
+    cnt = 0
+    import time
+    total = to_idx - from_idx
+    for scene in scene_list[from_idx:to_idx]:
+        start = time.time()
+        if os.path.exists(os.path.join(root, scene, "scan_points.txt")):
+            print(f"Scene {scene} has been processed")
+            cnt+=1
+            continue
+        save_scene_data(root, scene, cnt, total, file_type="npy")
+        cnt+=1
+        end = time.time()
+        print(f"Time cost: {end-start}")

runners/inferece_server.py

@@ -0,0 +1,109 @@
+import os
+import json
+import torch
+import numpy as np
+from flask import Flask, request, jsonify
+
+import PytorchBoot.namespace as namespace
+import PytorchBoot.stereotype as stereotype
+from PytorchBoot.factory import ComponentFactory
+
+from PytorchBoot.runners.runner import Runner
+from PytorchBoot.utils import Log
+
+from utils.pts import PtsUtil
+
+@stereotype.runner("inferencer")
+class InferencerServer(Runner):
+    def __init__(self, config_path):
+        super().__init__(config_path)
+    
+        ''' Web Server '''
+        self.app = Flask(__name__)
+        ''' Pipeline '''
+        self.pipeline_name = self.config[namespace.Stereotype.PIPELINE]
+        self.pipeline:torch.nn.Module = ComponentFactory.create(namespace.Stereotype.PIPELINE, self.pipeline_name)
+        self.pipeline = self.pipeline.to(self.device)
+            
+        ''' Experiment '''
+        self.load_experiment("nbv_evaluator")
+
+    def get_input_data(self, data):
+        input_data = {}
+        scanned_pts = data["scanned_pts"]
+        scanned_n_to_world_pose_9d = data["scanned_n_to_world_pose_9d"]
+        combined_scanned_views_pts = np.concatenate(scanned_pts, axis=0)
+        fps_downsampled_combined_scanned_pts, fps_idx = PtsUtil.fps_downsample_point_cloud(
+            combined_scanned_views_pts, self.pts_num, require_idx=True
+        )
+        combined_scanned_views_pts_mask = np.zeros(len(scanned_pts), dtype=np.uint8)
+        start_idx = 0
+        for i in range(len(scanned_pts)):
+            end_idx = start_idx + len(scanned_pts[i])
+            combined_scanned_views_pts_mask[start_idx:end_idx] = i
+            start_idx = end_idx
+
+        fps_downsampled_combined_scanned_pts_mask = combined_scanned_views_pts_mask[fps_idx]
+
+        input_data["scanned_pts_mask"] = np.asarray(fps_downsampled_combined_scanned_pts_mask, dtype=np.uint8)
+        input_data["scanned_n_to_world_pose_9d"] = np.asarray(scanned_n_to_world_pose_9d, dtype=np.float32)
+        input_data["combined_scanned_pts"] = np.asarray(fps_downsampled_combined_scanned_pts, dtype=np.float32)
+        return input_data
+    
+    def get_result(self, output_data):
+
+        estimated_delta_rot_9d = output_data["pred_pose_9d"]
+        result = {
+            "estimated_delta_rot_9d": estimated_delta_rot_9d.tolist()
+        }
+        return result
+    
+    def run(self):
+        Log.info("Loading from epoch {}.".format(self.current_epoch))
+    
+        @self.app.route("/inference", methods=["POST"])
+        def inference():
+            data = request.json
+            input_data = self.get_input_data(data)
+            output_data = self.pipeline.forward_test(input_data)
+            result = self.get_result(output_data)
+            return jsonify(result)
+        
+
+        self.app.run(host="0.0.0.0", port=5000)
+
+    def get_checkpoint_path(self, is_last=False):
+        return os.path.join(self.experiment_path, namespace.Direcotry.CHECKPOINT_DIR_NAME,
+                            "Epoch_{}.pth".format(
+                                self.current_epoch if self.current_epoch != -1 and not is_last else "last"))
+
+    def load_checkpoint(self, is_last=False):
+        self.load(self.get_checkpoint_path(is_last))
+        Log.success(f"Loaded checkpoint from {self.get_checkpoint_path(is_last)}")
+        if is_last:
+            checkpoint_root = os.path.join(self.experiment_path, namespace.Direcotry.CHECKPOINT_DIR_NAME)
+            meta_path = os.path.join(checkpoint_root, "meta.json")
+            if not os.path.exists(meta_path):
+                raise FileNotFoundError(
+                    "No checkpoint meta.json file in the experiment {}".format(self.experiments_config["name"]))
+            file_path = os.path.join(checkpoint_root, "meta.json")
+            with open(file_path, "r") as f:
+                meta = json.load(f)
+            self.current_epoch = meta["last_epoch"]
+            self.current_iter = meta["last_iter"]
+
+    def load_experiment(self, backup_name=None):
+        super().load_experiment(backup_name)
+        self.current_epoch = self.experiments_config["epoch"]
+        self.load_checkpoint(is_last=(self.current_epoch == -1))
+
+    def create_experiment(self, backup_name=None):
+        super().create_experiment(backup_name)
+       
+        
+    def load(self, path):
+        state_dict = torch.load(path)
+        self.pipeline.load_state_dict(state_dict)
+
+
+

runners/inferencer.py

@@ -20,7 +20,7 @@ from PytorchBoot.runners.runner import Runner
 from PytorchBoot.utils import Log
 from PytorchBoot.status import status_manager
 
-@stereotype.runner("inferencer", comment="not tested")
+@stereotype.runner("inferencer")
 class Inferencer(Runner):
     def __init__(self, config_path):
         super().__init__(config_path)
@@ -34,6 +34,7 @@ class Inferencer(Runner):
             
         ''' Experiment '''
         self.load_experiment("nbv_evaluator")
+        self.stat_result = {}
 
         ''' Test '''
         self.test_config = ConfigManager.get(namespace.Stereotype.RUNNER, namespace.Mode.TEST)
@@ -103,9 +104,9 @@ class Inferencer(Runner):
         input_data["scanned_pts"] = [data["first_pts"][0].to(self.device)]
         input_data["scanned_n_to_world_pose_9d"] = [data["first_to_world_9d"][0].to(self.device)]
         input_data["mode"] = namespace.Mode.TEST
+        input_data["combined_scanned_pts"] = data["combined_scanned_pts"]
         input_pts_N = input_data["scanned_pts"][0].shape[1]
         
-        
         first_frame_target_pts, _ = RenderUtil.render_pts(first_frame_to_world, scene_path, self.script_path, model_points_normals, voxel_threshold=voxel_threshold, filter_degree=filter_degree, nO_to_nL_pose=O_to_L_pose)
         scanned_view_pts = [first_frame_target_pts]
         last_pred_cr = self.compute_coverage_rate(scanned_view_pts, None, down_sampled_model_pts, threshold=voxel_threshold)
@@ -138,7 +139,7 @@ class Inferencer(Runner):
             
             print(pred_cr, last_pred_cr, " max: ", data["max_coverage_rate"])
             if pred_cr >= data["max_coverage_rate"]:
-                break
+                print("max coverage rate reached!")
             if pred_cr <= last_pred_cr + cr_increase_threshold:
                 retry += 1
                 retry_duplication_pose.append(pred_pose.cpu().numpy().tolist())
@@ -155,6 +156,11 @@ class Inferencer(Runner):
             
             input_data["scanned_pts"] = [torch.cat([input_data["scanned_pts"][0] , new_pts_tensor], dim=0)]
             input_data["scanned_n_to_world_pose_9d"] = [torch.cat([input_data["scanned_n_to_world_pose_9d"][0], pred_pose_9d], dim=0)]
+            combined_scanned_views_pts = np.concatenate(input_data["scanned_pts"][0].tolist(), axis=0)
+            voxel_downsampled_combined_scanned_pts_np = PtsUtil.voxel_downsample_point_cloud(combined_scanned_views_pts, 0.002)
+            random_downsampled_combined_scanned_pts_np = PtsUtil.random_downsample_point_cloud(voxel_downsampled_combined_scanned_pts_np, input_pts_N)
+            input_data["combined_scanned_pts"] = torch.tensor(random_downsampled_combined_scanned_pts_np, dtype=torch.float32).unsqueeze(0).to(self.device)
+
             
             last_pred_cr = pred_cr
 
@@ -173,6 +179,15 @@ class Inferencer(Runner):
             "retry_duplication_pose": retry_duplication_pose,
             "best_seq_len": data["best_seq_len"][0],
         }
+        self.stat_result[scene_name] = {
+            "max_coverage_rate": data["max_coverage_rate"][0],
+            "success_rate": max(pred_cr_seq)/ data["max_coverage_rate"][0],
+            "coverage_rate_seq": pred_cr_seq,
+            "pred_max_coverage_rate": max(pred_cr_seq),
+            "pred_seq_len": len(pred_cr_seq),
+        }
+        print('success rate: ', max(pred_cr_seq) / data["max_coverage_rate"][0])
+
         return result
 
     def compute_coverage_rate(self, scanned_view_pts, new_pts, model_pts, threshold=0.005):
@@ -191,6 +206,8 @@ class Inferencer(Runner):
             os.makedirs(dataset_dir)
         output_path = os.path.join(dataset_dir, f"{scene_name}.pkl")
         pickle.dump(output, open(output_path, "wb"))
+        with open(os.path.join(dataset_dir, "stat.json"), "w") as f:
+            json.dump(self.stat_result, f)
     
 
     def get_checkpoint_path(self, is_last=False):

runners/strategy_generator.py

@@ -22,43 +22,39 @@ class StrategyGenerator(Runner):
             "app_name": "generate_strategy",
             "runner_name": "strategy_generator"
         }
-        self.to_specified_dir = ConfigManager.get("runner", "generate", "to_specified_dir")
-        self.save_best_combined_pts = ConfigManager.get("runner", "generate", "save_best_combined_points")
-        self.save_mesh = ConfigManager.get("runner", "generate", "save_mesh")
-        self.load_pts = ConfigManager.get("runner", "generate", "load_points")
-        self.filter_degree = ConfigManager.get("runner", "generate", "filter_degree")
         self.overwrite = ConfigManager.get("runner", "generate", "overwrite")
-        self.save_pts = ConfigManager.get("runner","generate","save_points")
         self.seq_num = ConfigManager.get("runner","generate","seq_num")
+        self.overlap_area_threshold = ConfigManager.get("runner","generate","overlap_area_threshold")
+        self.compute_with_normal = ConfigManager.get("runner","generate","compute_with_normal")
+        self.scan_points_threshold = ConfigManager.get("runner","generate","scan_points_threshold")
         
 
         
     def run(self):
         dataset_name_list =  ConfigManager.get("runner", "generate", "dataset_list")
-        voxel_threshold, overlap_threshold = ConfigManager.get("runner","generate","voxel_threshold"), ConfigManager.get("runner","generate","overlap_threshold")
+        voxel_threshold = ConfigManager.get("runner","generate","voxel_threshold")
         for dataset_idx in range(len(dataset_name_list)):
             dataset_name = dataset_name_list[dataset_idx]
             status_manager.set_progress("generate_strategy", "strategy_generator", "dataset", dataset_idx, len(dataset_name_list))
             root_dir = ConfigManager.get("datasets", dataset_name, "root_dir")
-            model_dir = ConfigManager.get("datasets", dataset_name, "model_dir")
+            from_idx = ConfigManager.get("datasets",dataset_name,"from")
+            to_idx = ConfigManager.get("datasets",dataset_name,"to")
             scene_name_list = os.listdir(root_dir)
+            if to_idx == -1:
+                to_idx = len(scene_name_list)
             cnt = 0
-            total = len(scene_name_list)
-            for scene_name in scene_name_list:
+            total = len(scene_name_list[from_idx:to_idx])
+            Log.info(f"Processing Dataset: {dataset_name}, From: {from_idx}, To: {to_idx}")
+            for scene_name in scene_name_list[from_idx:to_idx]:
                 Log.info(f"({dataset_name})Processing [{cnt}/{total}]: {scene_name}")
                 status_manager.set_progress("generate_strategy", "strategy_generator", "scene", cnt, total)
-                diag = DataLoadUtil.get_bbox_diag(model_dir, scene_name)
-                voxel_threshold = diag*0.02
-                status_manager.set_status("generate_strategy", "strategy_generator", "voxel_threshold", voxel_threshold)
                 output_label_path = DataLoadUtil.get_label_path(root_dir, scene_name,0)
                 if os.path.exists(output_label_path) and not self.overwrite:
                     Log.info(f"Scene <{scene_name}> Already Exists, Skip")
                     cnt += 1
                     continue
-                try:
-                    self.generate_sequence(root_dir, model_dir, scene_name,voxel_threshold, overlap_threshold)
-                except Exception as e:
-                    Log.error(f"Scene <{scene_name}> Failed, Error: {e}")
+                
+                self.generate_sequence(root_dir, scene_name,voxel_threshold)
                 cnt += 1
             status_manager.set_progress("generate_strategy", "strategy_generator", "scene", total, total)
         status_manager.set_progress("generate_strategy", "strategy_generator", "dataset", len(dataset_name_list), len(dataset_name_list))
@@ -71,43 +67,61 @@ class StrategyGenerator(Runner):
     def load_experiment(self, backup_name=None):
         super().load_experiment(backup_name)
     
-    def generate_sequence(self, root, model_dir, scene_name, voxel_threshold, overlap_threshold):
+    def generate_sequence(self, root, scene_name, voxel_threshold):
         status_manager.set_status("generate_strategy", "strategy_generator", "scene", scene_name)
         frame_num = DataLoadUtil.get_scene_seq_length(root, scene_name)
+        
         model_points_normals = DataLoadUtil.load_points_normals(root, scene_name)
         model_pts = model_points_normals[:,:3]
-        down_sampled_model_pts = PtsUtil.voxel_downsample_point_cloud(model_pts, voxel_threshold)
-        
+        down_sampled_model_pts, idx = PtsUtil.voxel_downsample_point_cloud(model_pts, voxel_threshold, require_idx=True)
+        down_sampled_model_nrm = model_points_normals[idx, 3:]
         pts_list = []
-        
+        nrm_list = []
+        scan_points_indices_list = []
+        non_zero_cnt = 0
+
         for frame_idx in range(frame_num):
-            if self.load_pts and os.path.exists(os.path.join(root,scene_name, "pts", f"{frame_idx}.txt")):
-                sampled_point_cloud = np.loadtxt(os.path.join(root,scene_name, "pts", f"{frame_idx}.txt"))
-                status_manager.set_progress("generate_strategy", "strategy_generator", "loading frame", frame_idx, frame_num)
-                pts_list.append(sampled_point_cloud)
-                continue
-            else:
-                path = DataLoadUtil.get_path(root, scene_name, frame_idx)
-                cam_params = DataLoadUtil.load_cam_info(path, binocular=True)
-                status_manager.set_progress("generate_strategy", "strategy_generator", "loading frame", frame_idx, frame_num)
-                point_cloud = DataLoadUtil.get_target_point_cloud_world_from_path(path, binocular=True)
-                sampled_point_cloud = ReconstructionUtil.filter_points(point_cloud, model_points_normals, cam_pose=cam_params["cam_to_world"], voxel_size=voxel_threshold, theta=self.filter_degree)
-
-                if self.save_pts:
-                    pts_dir = os.path.join(root,scene_name, "pts")
-                    if not os.path.exists(pts_dir):
-                        os.makedirs(pts_dir)
-                    np.savetxt(os.path.join(pts_dir, f"{frame_idx}.txt"), sampled_point_cloud)
-                pts_list.append(sampled_point_cloud)
+            status_manager.set_progress("generate_strategy", "strategy_generator", "loading frame", frame_idx, frame_num)
+            pts_path = os.path.join(root,scene_name, "pts", f"{frame_idx}.npy")
+            nrm_path = os.path.join(root,scene_name, "nrm", f"{frame_idx}.npy")
+            idx_path = os.path.join(root,scene_name, "scan_points_indices", f"{frame_idx}.npy")
+            
+            pts = np.load(pts_path)
+            if self.compute_with_normal:
+                if pts.shape[0] == 0:
+                    nrm = np.zeros((0,3))
+                else:
+                    nrm = np.load(nrm_path)
+                nrm_list.append(nrm)
+            pts_list.append(pts)
+            indices = np.load(idx_path) 
+            scan_points_indices_list.append(indices)
+            if pts.shape[0] > 0:
+                non_zero_cnt += 1
         status_manager.set_progress("generate_strategy", "strategy_generator", "loading frame", frame_num, frame_num)
-
-        seq_num = min(self.seq_num, len(pts_list))
-        init_view_list = range(seq_num)
+        
+        seq_num = min(self.seq_num, non_zero_cnt)
+        init_view_list = []
+        idx = 0
+        while len(init_view_list) < seq_num and idx < len(pts_list):
+            if pts_list[idx].shape[0] > 50:
+                init_view_list.append(idx)
+            idx += 1
 
         seq_idx = 0
+        import time
         for init_view in init_view_list:
             status_manager.set_progress("generate_strategy", "strategy_generator", "computing sequence", seq_idx, len(init_view_list))
-            limited_useful_view, _, _ = ReconstructionUtil.compute_next_best_view_sequence_with_overlap(down_sampled_model_pts, pts_list,init_view=init_view, threshold=voxel_threshold, overlap_threshold=overlap_threshold, status_info=self.status_info)
+            start = time.time()
+
+            if not self.compute_with_normal:    
+                limited_useful_view, _, _ = ReconstructionUtil.compute_next_best_view_sequence(down_sampled_model_pts, pts_list, scan_points_indices_list = scan_points_indices_list,init_view=init_view, 
+                                                                                                            threshold=voxel_threshold, scan_points_threshold=self.scan_points_threshold, overlap_area_threshold=self.overlap_area_threshold, status_info=self.status_info)
+            else:
+                limited_useful_view, _, _ = ReconstructionUtil.compute_next_best_view_sequence_with_normal(down_sampled_model_pts, down_sampled_model_nrm, pts_list, nrm_list, scan_points_indices_list = scan_points_indices_list,init_view=init_view, 
+                                                                                                            threshold=voxel_threshold, scan_points_threshold=self.scan_points_threshold, overlap_area_threshold=self.overlap_area_threshold, status_info=self.status_info)
+            end = time.time()
+            print(f"Time: {end-start}")
             data_pairs = self.generate_data_pairs(limited_useful_view)
             seq_save_data = {
                 "data_pairs": data_pairs,
@@ -124,8 +138,6 @@ class StrategyGenerator(Runner):
             with open(output_label_path, 'w') as f:
                 json.dump(seq_save_data, f)
             seq_idx += 1
-        if self.save_mesh:
-            DataLoadUtil.save_target_mesh_at_world_space(root, model_dir, scene_name)
         status_manager.set_progress("generate_strategy", "strategy_generator", "computing sequence", len(init_view_list), len(init_view_list))
 
 

runners/view_generator.py

@@ -9,7 +9,8 @@ class ViewGenerator(Runner):
         self.config_path = config_path
         
     def run(self):
-        subprocess.run(['blender', '-b', '-P', '../blender/run_blender.py', '--', self.config_path])
+        result = subprocess.run(['/home/hofee/blender-4.0.2-linux-x64/blender', '-b', '-P', '../blender/run_blender.py', '--', self.config_path])
+        print()
 
     def create_experiment(self, backup_name=None):
         return super().create_experiment(backup_name)

utils/data_load.py

@@ -4,58 +4,79 @@ import json
 import cv2
 import trimesh
 import torch
+import OpenEXR
+import Imath
 from utils.pts import PtsUtil
 
+
 class DataLoadUtil:
-    TABLE_POSITION = np.asarray([0,0,0.8215])
+    TABLE_POSITION = np.asarray([0, 0, 0.8215])
+    
+    @staticmethod
+    def load_exr_image(file_path):
+        exr_file = OpenEXR.InputFile(file_path)
+        header = exr_file.header()
+        dw = header['dataWindow']
+        width = dw.max.x - dw.min.x + 1
+        height = dw.max.y - dw.min.y + 1
+        float_channels = ['R', 'G', 'B']
+        img_data = []
+        for channel in float_channels:
+            channel_data = exr_file.channel(channel)
+            img_data.append(np.frombuffer(channel_data, dtype=np.float16).reshape((height, width)))
+
+        # 将各通道组合成一个 (height, width, 3) 的 RGB 图像
+        img = np.stack(img_data, axis=-1)
+        return img
 
     @staticmethod
     def get_display_table_info(root, scene_name):
         scene_info = DataLoadUtil.load_scene_info(root, scene_name)
         display_table_info = scene_info["display_table"]
         return display_table_info
-    
+
     @staticmethod
     def get_display_table_top(root, scene_name):
-        display_table_height = DataLoadUtil.get_display_table_info(root, scene_name)["height"]
-        display_table_top = DataLoadUtil.TABLE_POSITION + np.asarray([0,0,display_table_height])
+        display_table_height = DataLoadUtil.get_display_table_info(root, scene_name)[
+            "height"
+        ]
+        display_table_top = DataLoadUtil.TABLE_POSITION + np.asarray(
+            [0, 0, display_table_height]
+        )
         return display_table_top
-    
+
     @staticmethod
     def get_path(root, scene_name, frame_idx):
         path = os.path.join(root, scene_name, f"{frame_idx}")
         return path
-    
+
     @staticmethod
     def get_label_num(root, scene_name):
-        label_dir = os.path.join(root,scene_name,"label")
+        label_dir = os.path.join(root, scene_name, "label")
+        if not os.path.exists(label_dir):
+            return 0
         return len(os.listdir(label_dir))
-    
+
     @staticmethod
     def get_label_path(root, scene_name, seq_idx):
-        label_dir = os.path.join(root,scene_name,"label")
+        label_dir = os.path.join(root, scene_name, "label")
         if not os.path.exists(label_dir):
             os.makedirs(label_dir)
-        path = os.path.join(label_dir,f"{seq_idx}.json")
+        path = os.path.join(label_dir, f"{seq_idx}.json")
         return path
-    
-    @staticmethod
-    def get_label_path_old(root, scene_name):
-        path = os.path.join(root,scene_name,"label.json")
-        return path
-    
+
     @staticmethod
     def get_scene_seq_length(root, scene_name):
         camera_params_path = os.path.join(root, scene_name, "camera_params")
         return len(os.listdir(camera_params_path))
-    
+
     @staticmethod
     def load_mesh_at(model_dir, object_name, world_object_pose):
         model_path = os.path.join(model_dir, object_name, "mesh.obj")
         mesh = trimesh.load(model_path)
         mesh.apply_transform(world_object_pose)
         return mesh
-    
+
     @staticmethod
     def get_bbox_diag(model_dir, object_name):
         model_path = os.path.join(model_dir, object_name, "mesh.obj")
@@ -63,55 +84,24 @@ class DataLoadUtil:
         bbox = mesh.bounding_box.extents
         diagonal_length = np.linalg.norm(bbox)
         return diagonal_length
-    
-    
-    @staticmethod
-    def save_mesh_at(model_dir, output_dir, object_name, scene_name, world_object_pose):
-        mesh = DataLoadUtil.load_mesh_at(model_dir, object_name, world_object_pose)
-        model_path = os.path.join(output_dir, scene_name, "world_mesh.obj")
-        mesh.export(model_path)
 
-    @staticmethod
-    def save_target_mesh_at_world_space(root, model_dir, scene_name, display_table_as_world_space_origin=True):
-        scene_info = DataLoadUtil.load_scene_info(root, scene_name)
-        target_name = scene_info["target_name"]
-        transformation = scene_info[target_name]
-        if display_table_as_world_space_origin:
-            location = transformation["location"] - DataLoadUtil.get_display_table_top(root, scene_name)
-        else:
-            location = transformation["location"]
-        rotation_euler = transformation["rotation_euler"]
-        pose_mat = trimesh.transformations.euler_matrix(*rotation_euler)
-        pose_mat[:3, 3] = location
-
-        mesh = DataLoadUtil.load_mesh_at(model_dir, target_name, pose_mat)
-        mesh_dir = os.path.join(root, scene_name, "mesh")
-        if not os.path.exists(mesh_dir):
-            os.makedirs(mesh_dir)
-        model_path = os.path.join(mesh_dir, "world_target_mesh.obj")
-        mesh.export(model_path)
-    
     @staticmethod
     def load_scene_info(root, scene_name):
         scene_info_path = os.path.join(root, scene_name, "scene_info.json")
         with open(scene_info_path, "r") as f:
             scene_info = json.load(f)
         return scene_info
-    
+
     @staticmethod
-    def load_target_object_pose(root, scene_name):
-        scene_info = DataLoadUtil.load_scene_info(root, scene_name)
-        target_name = scene_info["target_name"]
-        transformation = scene_info[target_name]
-        location = transformation["location"]
-        rotation_euler = transformation["rotation_euler"]
-        pose_mat = trimesh.transformations.euler_matrix(*rotation_euler)
-        pose_mat[:3, 3] = location
-        return pose_mat
-    
+    def load_target_pts_num_dict(root, scene_name):
+        target_pts_num_path = os.path.join(root, scene_name, "target_pts_num.json")
+        with open(target_pts_num_path, "r") as f:
+            target_pts_num_dict = json.load(f)
+        return target_pts_num_dict
+
     @staticmethod
-    def load_depth(path, min_depth=0.01,max_depth=5.0,binocular=False):
-        
+    def load_depth(path, min_depth=0.01, max_depth=5.0, binocular=False):
+
         def load_depth_from_real_path(real_path, min_depth, max_depth):
             depth = cv2.imread(real_path, cv2.IMREAD_UNCHANGED)
             depth = depth.astype(np.float32) / 65535.0
@@ -119,86 +109,152 @@ class DataLoadUtil:
             max_depth = max_depth
             depth_meters = min_depth + (max_depth - min_depth) * depth
             return depth_meters
-        
+
         if binocular:
-            depth_path_L = os.path.join(os.path.dirname(path), "depth", os.path.basename(path) + "_L.png")
-            depth_path_R = os.path.join(os.path.dirname(path), "depth", os.path.basename(path) + "_R.png")
-            depth_meters_L = load_depth_from_real_path(depth_path_L, min_depth, max_depth)
-            depth_meters_R = load_depth_from_real_path(depth_path_R, min_depth, max_depth)
+            depth_path_L = os.path.join(
+                os.path.dirname(path), "depth", os.path.basename(path) + "_L.png"
+            )
+            depth_path_R = os.path.join(
+                os.path.dirname(path), "depth", os.path.basename(path) + "_R.png"
+            )
+            depth_meters_L = load_depth_from_real_path(
+                depth_path_L, min_depth, max_depth
+            )
+            depth_meters_R = load_depth_from_real_path(
+                depth_path_R, min_depth, max_depth
+            )
             return depth_meters_L, depth_meters_R
         else:
-            depth_path = os.path.join(os.path.dirname(path), "depth", os.path.basename(path) + ".png")
+            depth_path = os.path.join(
+                os.path.dirname(path), "depth", os.path.basename(path) + ".png"
+            )
             depth_meters = load_depth_from_real_path(depth_path, min_depth, max_depth)
             return depth_meters
-        
+
     @staticmethod
-    def load_seg(path, binocular=False):
-        if binocular:
+    def load_seg(path, binocular=False, left_only=False):
+        if binocular and not left_only:
+
             def clean_mask(mask_image):
-                green = [0, 255, 0, 255]
-                red = [255, 0, 0, 255]
+                green = [0, 255, 0]
+                red = [255, 0, 0]
                 threshold = 2
-                mask_image = np.where(np.abs(mask_image - green) <= threshold, green, mask_image)
-                mask_image = np.where(np.abs(mask_image - red) <= threshold, red, mask_image)
+                mask_image = np.where(
+                    np.abs(mask_image - green) <= threshold, green, mask_image
+                )
+                mask_image = np.where(
+                    np.abs(mask_image - red) <= threshold, red, mask_image
+                )
                 return mask_image
-            mask_path_L = os.path.join(os.path.dirname(path), "mask", os.path.basename(path) + "_L.png")
+
+            mask_path_L = os.path.join(
+                os.path.dirname(path), "mask", os.path.basename(path) + "_L.png"
+            )
             mask_image_L = clean_mask(cv2.imread(mask_path_L, cv2.IMREAD_UNCHANGED))
-            mask_path_R = os.path.join(os.path.dirname(path), "mask", os.path.basename(path) + "_R.png")
+            mask_path_R = os.path.join(
+                os.path.dirname(path), "mask", os.path.basename(path) + "_R.png"
+            )
             mask_image_R = clean_mask(cv2.imread(mask_path_R, cv2.IMREAD_UNCHANGED))
             return mask_image_L, mask_image_R
         else:
-            mask_path = os.path.join(os.path.dirname(path), "mask", os.path.basename(path) + ".png")
-            mask_image = cv2.imread(mask_path, cv2.IMREAD_GRAYSCALE)
+            if binocular and left_only:
+                mask_path = os.path.join(
+                    os.path.dirname(path), "mask", os.path.basename(path) + "_L.png"
+                )
+            else:
+                mask_path = os.path.join(
+                    os.path.dirname(path), "mask", os.path.basename(path) + ".png"
+                )
+            mask_image = cv2.imread(mask_path, cv2.IMREAD_UNCHANGED)
             return mask_image
-    
+        
+    @staticmethod
+    def load_normal(path, binocular=False, left_only=False, file_type="exr"):
+        if binocular and not left_only:
+            normal_path_L = os.path.join(
+                os.path.dirname(path), "normal", os.path.basename(path) + f"_L.{file_type}"
+            )
+            normal_image_L = DataLoadUtil.load_exr_image(normal_path_L)
+            
+            normal_path_R = os.path.join(
+                os.path.dirname(path), "normal", os.path.basename(path) + f"_R.{file_type}"
+            )
+            normal_image_R = DataLoadUtil.load_exr_image(normal_path_R)
+            normalized_normal_image_L = normal_image_L * 2.0 - 1.0
+            normalized_normal_image_R = normal_image_R * 2.0 - 1.0
+            return normalized_normal_image_L, normalized_normal_image_R
+        else:
+            if binocular and left_only:
+                normal_path = os.path.join(
+                    os.path.dirname(path), "normal", os.path.basename(path) + f"_L.{file_type}"
+                )
+            else:
+                normal_path = os.path.join(
+                    os.path.dirname(path), "normal", os.path.basename(path) + f".{file_type}"
+                )
+            normal_image = DataLoadUtil.load_exr_image(normal_path)
+            normalized_normal_image = normal_image * 2.0 - 1.0
+            return normalized_normal_image
+
     @staticmethod
     def load_label(path):
-        with open(path, 'r') as f:
+        with open(path, "r") as f:
             label_data = json.load(f)
         return label_data
-    
+
     @staticmethod
-    def load_rgb(path):
-        rgb_path = os.path.join(os.path.dirname(path), "rgb", os.path.basename(path) + ".png")
-        rgb_image = cv2.imread(rgb_path, cv2.IMREAD_COLOR)
-        return rgb_image
-    
-    @staticmethod
-    def load_from_preprocessed_pts(path):
-        npy_path = os.path.join(os.path.dirname(path), "points", os.path.basename(path) + ".npy")
-        pts = np.load(npy_path)
+    def load_from_preprocessed_pts(path, file_type="npy"):
+        npy_path = os.path.join(
+            os.path.dirname(path), "pts", os.path.basename(path) + "." + file_type
+        )
+        if file_type == "txt":
+            pts = np.loadtxt(npy_path)
+        else:
+            pts = np.load(npy_path)
         return pts
+    
+    @staticmethod
+    def load_from_preprocessed_nrm(path, file_type="npy"):
+        npy_path = os.path.join(
+            os.path.dirname(path), "nrm", os.path.basename(path) + "." + file_type
+        )
+        if file_type == "txt":
+            nrm = np.loadtxt(npy_path)
+        else:
+            nrm = np.load(npy_path)
+        return nrm
 
     @staticmethod
     def cam_pose_transformation(cam_pose_before):
-        offset = np.asarray([
-            [1, 0, 0, 0],
-            [0, -1, 0, 0],
-            [0, 0, -1, 0],
-            [0, 0, 0, 1]])  
-        cam_pose_after =  cam_pose_before @ offset
+        offset = np.asarray([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]])
+        cam_pose_after = cam_pose_before @ offset
         return cam_pose_after
-    
+
     @staticmethod
     def load_cam_info(path, binocular=False, display_table_as_world_space_origin=True):
         scene_dir = os.path.dirname(path)
         root_dir = os.path.dirname(scene_dir)
         scene_name = os.path.basename(scene_dir)
-        camera_params_path = os.path.join(os.path.dirname(path), "camera_params", os.path.basename(path) + ".json")
-        with open(camera_params_path, 'r') as f:
+        camera_params_path = os.path.join(
+            os.path.dirname(path), "camera_params", os.path.basename(path) + ".json"
+        )
+        with open(camera_params_path, "r") as f:
             label_data = json.load(f)
         cam_to_world = np.asarray(label_data["extrinsic"])
         cam_to_world = DataLoadUtil.cam_pose_transformation(cam_to_world)
-        world_to_display_table = np.eye(4)
-        world_to_display_table[:3, 3] = - DataLoadUtil.get_display_table_top(root_dir, scene_name)
+        
         if display_table_as_world_space_origin:
+            world_to_display_table = np.eye(4)
+            world_to_display_table[:3, 3] = -DataLoadUtil.get_display_table_top(
+                root_dir, scene_name
+            )
             cam_to_world = np.dot(world_to_display_table, cam_to_world)
         cam_intrinsic = np.asarray(label_data["intrinsic"])
         cam_info = {
             "cam_to_world": cam_to_world,
             "cam_intrinsic": cam_intrinsic,
             "far_plane": label_data["far_plane"],
-            "near_plane": label_data["near_plane"]
+            "near_plane": label_data["near_plane"],
         }
         if binocular:
             cam_to_world_R = np.asarray(label_data["extrinsic_R"])
@@ -211,104 +267,127 @@ class DataLoadUtil:
             cam_info["cam_to_world_O"] = cam_to_world_O
             cam_info["cam_to_world_R"] = cam_to_world_R
         return cam_info
-    
+
     @staticmethod
-    def get_real_cam_O_from_cam_L(cam_L, cam_O_to_cam_L, scene_path, display_table_as_world_space_origin=True):
+    def get_real_cam_O_from_cam_L(
+        cam_L, cam_O_to_cam_L, scene_path, display_table_as_world_space_origin=True
+    ):
         root_dir = os.path.dirname(scene_path)
         scene_name = os.path.basename(scene_path)
         if isinstance(cam_L, torch.Tensor):
             cam_L = cam_L.cpu().numpy()
-        nO_to_display_table_pose = cam_L  @ cam_O_to_cam_L
+        nO_to_display_table_pose = cam_L @ cam_O_to_cam_L
         if display_table_as_world_space_origin:
             display_table_to_world = np.eye(4)
-            display_table_to_world[:3, 3] = DataLoadUtil.get_display_table_top(root_dir, scene_name)
+            display_table_to_world[:3, 3] = DataLoadUtil.get_display_table_top(
+                root_dir, scene_name
+            )
             nO_to_world_pose = np.dot(display_table_to_world, nO_to_display_table_pose)
         nO_to_world_pose = DataLoadUtil.cam_pose_transformation(nO_to_world_pose)
         return nO_to_world_pose
-    
+
     @staticmethod
-    def get_target_point_cloud(depth, cam_intrinsic, cam_extrinsic, mask, target_mask_label=(0,255,0,255)):
+    def get_target_point_cloud(
+        depth, cam_intrinsic, cam_extrinsic, mask, target_mask_label=(0, 255, 0, 255), require_full_points=False
+    ):
         h, w = depth.shape
-        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing='xy')
-        
+        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing="xy")
+
         z = depth
         x = (i - cam_intrinsic[0, 2]) * z / cam_intrinsic[0, 0]
         y = (j - cam_intrinsic[1, 2]) * z / cam_intrinsic[1, 1]
-        
-        points_camera = np.stack((x, y, z), axis=-1).reshape(-1, 3)
-        mask = mask.reshape(-1,4)
 
-        target_mask = (mask == target_mask_label).all(axis=-1)  
-        
+        points_camera = np.stack((x, y, z), axis=-1).reshape(-1, 3)
+        mask = mask.reshape(-1, 4)
+
+        target_mask = (mask == target_mask_label).all(axis=-1)
+
         target_points_camera = points_camera[target_mask]
-        target_points_camera_aug = np.concatenate([target_points_camera, np.ones((target_points_camera.shape[0], 1))], axis=-1)
-        
+        target_points_camera_aug = np.concatenate(
+            [target_points_camera, np.ones((target_points_camera.shape[0], 1))], axis=-1
+        )
+
         target_points_world = np.dot(cam_extrinsic, target_points_camera_aug.T).T[:, :3]
-        return {
+        data = {
             "points_world": target_points_world,
-            "points_camera": target_points_camera
+            "points_camera": target_points_camera,
         }
-    
+        return data
+
     @staticmethod
     def get_point_cloud(depth, cam_intrinsic, cam_extrinsic):
         h, w = depth.shape
-        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing='xy')
-        
+        i, j = np.meshgrid(np.arange(w), np.arange(h), indexing="xy")
+
         z = depth
         x = (i - cam_intrinsic[0, 2]) * z / cam_intrinsic[0, 0]
         y = (j - cam_intrinsic[1, 2]) * z / cam_intrinsic[1, 1]
-        
+
         points_camera = np.stack((x, y, z), axis=-1).reshape(-1, 3)
-        points_camera_aug = np.concatenate([points_camera, np.ones((points_camera.shape[0], 1))], axis=-1)
-        
+        points_camera_aug = np.concatenate(
+            [points_camera, np.ones((points_camera.shape[0], 1))], axis=-1
+        )
+
         points_world = np.dot(cam_extrinsic, points_camera_aug.T).T[:, :3]
-        return {
-            "points_world": points_world,
-            "points_camera": points_camera
-        }
-        
+        return {"points_world": points_world, "points_camera": points_camera}
+
     @staticmethod
-    def get_target_point_cloud_world_from_path(path, binocular=False, random_downsample_N=65536, voxel_size = 0.005, target_mask_label=(0,255,0,255)):
+    def get_target_point_cloud_world_from_path(
+        path,
+        binocular=False,
+        random_downsample_N=65536,
+        voxel_size=0.005,
+        target_mask_label=(0, 255, 0, 255),
+        display_table_mask_label=(0, 0, 255, 255),
+        get_display_table_pts=False,
+        require_normal=False,
+    ):
         cam_info = DataLoadUtil.load_cam_info(path, binocular=binocular)
         if binocular:
-            depth_L, depth_R = DataLoadUtil.load_depth(path, cam_info['near_plane'], cam_info['far_plane'], binocular=True)
+            depth_L, depth_R = DataLoadUtil.load_depth(
+                path, cam_info["near_plane"], cam_info["far_plane"], binocular=True
+            )
             mask_L, mask_R = DataLoadUtil.load_seg(path, binocular=True)
-            point_cloud_L = DataLoadUtil.get_target_point_cloud(depth_L, cam_info['cam_intrinsic'], cam_info['cam_to_world'], mask_L, target_mask_label)['points_world']
-            point_cloud_R = DataLoadUtil.get_target_point_cloud(depth_R, cam_info['cam_intrinsic'], cam_info['cam_to_world_R'], mask_R, target_mask_label)['points_world']
-            point_cloud_L = PtsUtil.random_downsample_point_cloud(point_cloud_L, random_downsample_N)
-            point_cloud_R = PtsUtil.random_downsample_point_cloud(point_cloud_R, random_downsample_N)
-            overlap_points = DataLoadUtil.get_overlapping_points(point_cloud_L, point_cloud_R, voxel_size)
+            point_cloud_L = DataLoadUtil.get_target_point_cloud(
+                depth_L,
+                cam_info["cam_intrinsic"],
+                cam_info["cam_to_world"],
+                mask_L,
+                target_mask_label,
+            )["points_world"]
+            point_cloud_R = DataLoadUtil.get_target_point_cloud(
+                depth_R,
+                cam_info["cam_intrinsic"],
+                cam_info["cam_to_world_R"],
+                mask_R,
+                target_mask_label,
+            )["points_world"]
+            point_cloud_L = PtsUtil.random_downsample_point_cloud(
+                point_cloud_L, random_downsample_N
+            )
+            point_cloud_R = PtsUtil.random_downsample_point_cloud(
+                point_cloud_R, random_downsample_N
+            )
+            overlap_points = PtsUtil.get_overlapping_points(
+                point_cloud_L, point_cloud_R, voxel_size
+            )
             return overlap_points
         else:
-            depth = DataLoadUtil.load_depth(path, cam_info['near_plane'], cam_info['far_plane'])
+            depth = DataLoadUtil.load_depth(
+                path, cam_info["near_plane"], cam_info["far_plane"]
+            )
             mask = DataLoadUtil.load_seg(path)
-            point_cloud = DataLoadUtil.get_target_point_cloud(depth, cam_info['cam_intrinsic'], cam_info['cam_to_world'], mask)['points_world']
+            point_cloud = DataLoadUtil.get_target_point_cloud(
+                depth, cam_info["cam_intrinsic"], cam_info["cam_to_world"], mask
+            )["points_world"]
             return point_cloud
-            
-    
-    @staticmethod
-    def voxelize_points(points, voxel_size):
- 
-        voxel_indices = np.floor(points / voxel_size).astype(np.int32)
-        unique_voxels = np.unique(voxel_indices, axis=0, return_inverse=True)
-        return unique_voxels
-    
-    @staticmethod
-    def get_overlapping_points(point_cloud_L, point_cloud_R, voxel_size=0.005):
-        voxels_L, indices_L = DataLoadUtil.voxelize_points(point_cloud_L, voxel_size)
-        voxels_R, _ = DataLoadUtil.voxelize_points(point_cloud_R, voxel_size)
 
-        voxel_indices_L = voxels_L.view([('', voxels_L.dtype)]*3)
-        voxel_indices_R = voxels_R.view([('', voxels_R.dtype)]*3)
-        overlapping_voxels = np.intersect1d(voxel_indices_L, voxel_indices_R)
-        mask_L = np.isin(indices_L, np.where(np.isin(voxel_indices_L, overlapping_voxels))[0])
-        overlapping_points = point_cloud_L[mask_L]
-        return overlapping_points
-    
     @staticmethod
     def load_points_normals(root, scene_name, display_table_as_world_space_origin=True):
         points_path = os.path.join(root, scene_name, "points_and_normals.txt")
         points_normals = np.loadtxt(points_path)
         if display_table_as_world_space_origin:
-            points_normals[:,:3] = points_normals[:,:3] - DataLoadUtil.get_display_table_top(root, scene_name)
-        return points_normals
+            points_normals[:, :3] = points_normals[
+                :, :3
+            ] - DataLoadUtil.get_display_table_top(root, scene_name)
+        return points_normals

utils/pts.py

@@ -5,12 +5,83 @@ import torch
 class PtsUtil:
 
     @staticmethod
-    def voxel_downsample_point_cloud(point_cloud, voxel_size=0.005):
-        o3d_pc = o3d.geometry.PointCloud()
-        o3d_pc.points = o3d.utility.Vector3dVector(point_cloud)
-        downsampled_pc = o3d_pc.voxel_down_sample(voxel_size)
-        return np.asarray(downsampled_pc.points)
+    def voxel_downsample_point_cloud(point_cloud, voxel_size=0.005, require_idx=False):
+        voxel_indices = np.floor(point_cloud / voxel_size).astype(np.int32)
+        if require_idx:
+            _, inverse, counts = np.unique(voxel_indices, axis=0, return_inverse=True, return_counts=True)
+            idx_sort = np.argsort(inverse)
+            idx_unique = idx_sort[np.cumsum(counts)-counts]
+            downsampled_points = point_cloud[idx_unique]
+            return downsampled_points, idx_unique
+        else:
+            import ipdb; ipdb.set_trace()
+            unique_voxels = np.unique(voxel_indices, axis=0, return_inverse=False)
+            return unique_voxels*voxel_size
+
+    @staticmethod
+    def voxel_downsample_point_cloud_o3d(point_cloud, voxel_size=0.005):
+        pcd = o3d.geometry.PointCloud()
+        pcd.points = o3d.utility.Vector3dVector(point_cloud)
+        pcd = pcd.voxel_down_sample(voxel_size)
+        return np.asarray(pcd.points)
     
+    @staticmethod
+    def voxel_downsample_point_cloud_and_trace_o3d(point_cloud, voxel_size=0.005):
+        pcd = o3d.geometry.PointCloud()
+        pcd.points = o3d.utility.Vector3dVector(point_cloud)
+        max_bound = pcd.get_max_bound()
+        min_bound = pcd.get_min_bound()
+        pcd = pcd.voxel_down_sample_and_trace(voxel_size, max_bound, min_bound, True)
+        
+        return np.asarray(pcd.points)
+       
+    @staticmethod
+    def random_downsample_point_cloud(point_cloud, num_points, require_idx=False, replace=True):
+        if point_cloud.shape[0] == 0:
+            if require_idx:
+                return point_cloud, np.array([])
+            return point_cloud
+        if not replace and num_points > len(point_cloud):
+            if require_idx:
+                return point_cloud, np.arange(len(point_cloud))
+            return point_cloud
+        idx = np.random.choice(len(point_cloud), num_points, replace=replace)
+        if require_idx:
+            return point_cloud[idx], idx
+        return point_cloud[idx]
+    
+    @staticmethod
+    def fps_downsample_point_cloud(point_cloud, num_points, require_idx=False):
+        N = point_cloud.shape[0]
+        mask = np.zeros(N, dtype=bool)
+
+        sampled_indices = np.zeros(num_points, dtype=int)
+        sampled_indices[0] = np.random.randint(0, N)
+        distances = np.linalg.norm(point_cloud - point_cloud[sampled_indices[0]], axis=1)
+        for i in range(1, num_points):
+            farthest_index = np.argmax(distances)
+            sampled_indices[i] = farthest_index
+            mask[farthest_index] = True
+
+            new_distances = np.linalg.norm(point_cloud - point_cloud[farthest_index], axis=1)
+            distances = np.minimum(distances, new_distances)
+        
+        sampled_points = point_cloud[sampled_indices]
+        if require_idx:
+            return sampled_points, sampled_indices
+        return sampled_points
+        
+    @staticmethod
+    def random_downsample_point_cloud_tensor(point_cloud, num_points):
+        idx = torch.randint(0, len(point_cloud), (num_points,))
+        return point_cloud[idx]
+    
+    @staticmethod
+    def voxelize_points(points, voxel_size):
+        voxel_indices = np.floor(points / voxel_size).astype(np.int32)
+        unique_voxels = np.unique(voxel_indices, axis=0, return_inverse=True)
+        return unique_voxels
+
     @staticmethod
     def transform_point_cloud(points, pose_mat):
         points_h = np.concatenate([points, np.ones((points.shape[0], 1))], axis=1)
@@ -18,11 +89,40 @@ class PtsUtil:
         return points_h[:, :3]
     
     @staticmethod
-    def random_downsample_point_cloud(point_cloud, num_points):
-        idx = np.random.choice(len(point_cloud), num_points, replace=True)
-        return point_cloud[idx]
+    def get_overlapping_points(point_cloud_L, point_cloud_R, voxel_size=0.005, require_idx=False):
+        voxels_L, indices_L = PtsUtil.voxelize_points(point_cloud_L, voxel_size)
+        voxels_R, _ = PtsUtil.voxelize_points(point_cloud_R, voxel_size)
+
+        voxel_indices_L = voxels_L.view([("", voxels_L.dtype)] * 3)
+        voxel_indices_R = voxels_R.view([("", voxels_R.dtype)] * 3)
+        overlapping_voxels = np.intersect1d(voxel_indices_L, voxel_indices_R)
+        mask_L = np.isin(
+            indices_L, np.where(np.isin(voxel_indices_L, overlapping_voxels))[0]
+        )
+        overlapping_points = point_cloud_L[mask_L]
+        if require_idx:
+            return overlapping_points, mask_L
+        return overlapping_points
     
     @staticmethod
-    def random_downsample_point_cloud_tensor(point_cloud, num_points):
-        idx = torch.randint(0, len(point_cloud), (num_points,))
-        return point_cloud[idx]
+    def filter_points(points, normals, cam_pose, theta_limit=45, z_range=(0.2, 0.45)):
+        
+        """ filter with normal """ 
+        normals_normalized = normals / np.linalg.norm(normals, axis=1, keepdims=True)
+        cos_theta = np.dot(normals_normalized, np.array([0, 0, 1]))
+        theta = np.arccos(cos_theta) * 180 / np.pi
+        idx = theta < theta_limit
+        filtered_sampled_points = points[idx]
+        filtered_normals = normals[idx]
+        
+        """ filter with z range """
+        points_cam = PtsUtil.transform_point_cloud(filtered_sampled_points, np.linalg.inv(cam_pose))
+        idx = (points_cam[:, 2] > z_range[0]) & (points_cam[:, 2] < z_range[1])
+        z_filtered_points = filtered_sampled_points[idx]
+        z_filtered_normals = filtered_normals[idx]
+        return z_filtered_points[:, :3], z_filtered_normals
+    
+    @staticmethod
+    def point_to_hash(point, voxel_size):
+        return tuple(np.floor(point / voxel_size).astype(int))
+    

utils/reconstruction.py

@@ -3,88 +3,126 @@ from scipy.spatial import cKDTree
 from utils.pts import PtsUtil
 
 class ReconstructionUtil:
-    
+
     @staticmethod
     def compute_coverage_rate(target_point_cloud, combined_point_cloud, threshold=0.01):
         kdtree = cKDTree(combined_point_cloud)
         distances, _ = kdtree.query(target_point_cloud)
-        covered_points = np.sum(distances < threshold)
-        coverage_rate = covered_points / target_point_cloud.shape[0]
-        return coverage_rate
+        covered_points_num = np.sum(distances < threshold*2)
+        coverage_rate = covered_points_num / target_point_cloud.shape[0]
+        return coverage_rate, covered_points_num
     
     @staticmethod
-    def compute_overlap_rate(new_point_cloud, combined_point_cloud, threshold=0.01):
+    def compute_coverage_rate_with_normal(target_point_cloud, combined_point_cloud, target_normal, combined_normal, threshold=0.01, normal_threshold=0.1):
+        kdtree = cKDTree(combined_point_cloud)
+        distances, indices = kdtree.query(target_point_cloud)
+        is_covered_by_distance = distances < threshold*2
+        normal_dots = np.einsum('ij,ij->i', target_normal, combined_normal[indices])
+        is_covered_by_normal = normal_dots > normal_threshold
+
+        pts_nrm_target = np.hstack([target_point_cloud, target_normal])
+        np.savetxt("pts_nrm_target.txt",  pts_nrm_target)
+        pts_nrm_combined = np.hstack([combined_point_cloud, combined_normal])
+        np.savetxt("pts_nrm_combined.txt", pts_nrm_combined)
+        import ipdb; ipdb.set_trace()
+        covered_points_num = np.sum(is_covered_by_distance & is_covered_by_normal)
+        coverage_rate = covered_points_num / target_point_cloud.shape[0]
+
+        return coverage_rate, covered_points_num
+        
+    
+    @staticmethod
+    def check_overlap(new_point_cloud, combined_point_cloud, overlap_area_threshold=25, voxel_size=0.01):
         kdtree = cKDTree(combined_point_cloud)
         distances, _ = kdtree.query(new_point_cloud)
-        overlapping_points = np.sum(distances < threshold)
-        overlap_rate = overlapping_points / new_point_cloud.shape[0]
-        return overlap_rate
-    
-    @staticmethod
-    def combine_point_with_view_sequence(point_list, view_sequence):
-        selected_views = []
-        for view_index, _ in view_sequence:
-            selected_views.append(point_list[view_index])
-        return np.vstack(selected_views)
-    
-    @staticmethod 
-    def compute_next_view_coverage_list(views, combined_point_cloud, target_point_cloud, threshold=0.01):
-        best_view = None
-        best_coverage_increase = -1
-        current_coverage = ReconstructionUtil.compute_coverage_rate(target_point_cloud, combined_point_cloud, threshold)
-        
-        for view_index, view in enumerate(views):
-            candidate_views = combined_point_cloud + [view]
-            down_sampled_combined_point_cloud = PtsUtil.voxel_downsample_point_cloud(candidate_views, threshold)
-            new_coverage = ReconstructionUtil.compute_coverage_rate(target_point_cloud, down_sampled_combined_point_cloud, threshold)
-            coverage_increase = new_coverage - current_coverage
-            if coverage_increase > best_coverage_increase:
-                best_coverage_increase = coverage_increase
-                best_view = view_index
-        return best_view, best_coverage_increase
+        overlapping_points = np.sum(distances < voxel_size*2)
+        cm = 0.01
+        voxel_size_cm = voxel_size / cm
+        overlap_area = overlapping_points * voxel_size_cm * voxel_size_cm
+        return overlap_area > overlap_area_threshold
 
+
+    @staticmethod
+    def get_new_added_points(old_combined_pts, new_pts, threshold=0.005):
+        if old_combined_pts.size == 0:
+            return new_pts
+        if new_pts.size == 0:
+            return np.array([])
+
+        tree = cKDTree(old_combined_pts)
+        distances, _ = tree.query(new_pts, k=1)
+        new_added_points = new_pts[distances > threshold]
+        return new_added_points
     
     @staticmethod
-    def compute_next_best_view_sequence_with_overlap(target_point_cloud, point_cloud_list,threshold=0.01, overlap_threshold=0.3, init_view = 0, status_info=None):
-        selected_views = [point_cloud_list[init_view]]
-        combined_point_cloud = np.vstack(selected_views)
-        down_sampled_combined_point_cloud = PtsUtil.voxel_downsample_point_cloud(combined_point_cloud,threshold)
-        new_coverage = ReconstructionUtil.compute_coverage_rate(target_point_cloud, down_sampled_combined_point_cloud, threshold)
+    def compute_next_best_view_sequence(target_point_cloud, point_cloud_list, scan_points_indices_list, threshold=0.01, overlap_area_threshold=25, init_view = 0, scan_points_threshold=5, status_info=None):
+        selected_views = [init_view]
+        combined_point_cloud = point_cloud_list[init_view]
+        history_indices = [scan_points_indices_list[init_view]]
+        
+        max_rec_pts = np.vstack(point_cloud_list)
+        downsampled_max_rec_pts = PtsUtil.voxel_downsample_point_cloud(max_rec_pts, threshold)
+        combined_point_cloud = PtsUtil.voxel_downsample_point_cloud(combined_point_cloud, threshold)
+        max_rec_pts_num = downsampled_max_rec_pts.shape[0]
+        max_real_rec_pts_coverage, _ = ReconstructionUtil.compute_coverage_rate(target_point_cloud, downsampled_max_rec_pts, threshold)
+        
+        new_coverage, new_covered_num = ReconstructionUtil.compute_coverage_rate(downsampled_max_rec_pts, combined_point_cloud, threshold)
         current_coverage = new_coverage
+        current_covered_num = new_covered_num
+        
         remaining_views = list(range(len(point_cloud_list)))
         view_sequence = [(init_view, current_coverage)]
         cnt_processed_view = 0
         remaining_views.remove(init_view)
-
+        curr_rec_pts_num = combined_point_cloud.shape[0]
+        drop_output_ratio = 0.4
+        
+        import time
         while remaining_views:
             best_view = None
             best_coverage_increase = -1
+            best_combined_point_cloud = None
+            best_covered_num = 0
         
             for view_index in remaining_views:
-                
+                if np.random.rand() < drop_output_ratio:
+                    continue
+                if point_cloud_list[view_index].shape[0] == 0:
+                    continue
                 if selected_views:
-                    combined_old_point_cloud = np.vstack(selected_views)
-                    down_sampled_old_point_cloud = PtsUtil.voxel_downsample_point_cloud(combined_old_point_cloud,threshold)
-                    down_sampled_new_view_point_cloud = PtsUtil.voxel_downsample_point_cloud(point_cloud_list[view_index],threshold)
-                    overlap_rate = ReconstructionUtil.compute_overlap_rate(down_sampled_new_view_point_cloud,down_sampled_old_point_cloud, threshold)
-                    if overlap_rate < overlap_threshold:
+                    new_scan_points_indices = scan_points_indices_list[view_index]
+                    if not ReconstructionUtil.check_scan_points_overlap(history_indices, new_scan_points_indices, scan_points_threshold):
+                        curr_overlap_area_threshold = overlap_area_threshold
+                    else:
+                        curr_overlap_area_threshold = overlap_area_threshold * 0.5
+                        
+                    if not ReconstructionUtil.check_overlap(point_cloud_list[view_index], combined_point_cloud, overlap_area_threshold = curr_overlap_area_threshold, voxel_size=threshold):
                         continue
-                    
-                candidate_views = selected_views + [point_cloud_list[view_index]]
-                combined_point_cloud = np.vstack(candidate_views)
-                down_sampled_combined_point_cloud = PtsUtil.voxel_downsample_point_cloud(combined_point_cloud,threshold)
-                new_coverage = ReconstructionUtil.compute_coverage_rate(target_point_cloud, down_sampled_combined_point_cloud, threshold)
+                
+                new_combined_point_cloud = np.vstack([combined_point_cloud, point_cloud_list[view_index]])
+                new_downsampled_combined_point_cloud = PtsUtil.voxel_downsample_point_cloud(new_combined_point_cloud,threshold)
+                new_coverage, new_covered_num = ReconstructionUtil.compute_coverage_rate(downsampled_max_rec_pts, new_downsampled_combined_point_cloud, threshold)
                 coverage_increase = new_coverage - current_coverage
                 if coverage_increase > best_coverage_increase:
                     best_coverage_increase = coverage_increase
                     best_view = view_index
+                    best_covered_num = new_covered_num
+                    best_combined_point_cloud = new_downsampled_combined_point_cloud
                     
             
             if best_view is not None:
-                if best_coverage_increase <=3e-3:
+                if best_coverage_increase <=1e-3 or best_covered_num - current_covered_num <= 5:
                     break
-                selected_views.append(point_cloud_list[best_view])
+                
+                selected_views.append(best_view)
+                best_rec_pts_num = best_combined_point_cloud.shape[0]
+                print(f"Current rec pts num: {curr_rec_pts_num}, Best rec pts num: {best_rec_pts_num}, Best cover pts: {best_covered_num}, Max rec pts num: {max_rec_pts_num}")
+                print(f"Current coverage: {current_coverage+best_coverage_increase}, Best coverage increase: {best_coverage_increase}, Max Real coverage: {max_real_rec_pts_coverage}")
+                current_covered_num = best_covered_num
+                curr_rec_pts_num = best_rec_pts_num
+                combined_point_cloud = best_combined_point_cloud
                 remaining_views.remove(best_view)
+                history_indices.append(scan_points_indices_list[best_view])
                 current_coverage += best_coverage_increase
                 cnt_processed_view += 1
                 if status_info is not None:
@@ -103,21 +141,125 @@ class ReconstructionUtil:
             app_name = status_info["app_name"]
             runner_name = status_info["runner_name"]
             sm.set_progress(app_name, runner_name, "processed view", len(point_cloud_list), len(point_cloud_list))
-        return view_sequence, remaining_views, down_sampled_combined_point_cloud
+        return view_sequence, remaining_views, combined_point_cloud
     
     @staticmethod
-    def filter_points(points, points_normals, cam_pose,  voxel_size=0.005, theta=45):
-        sampled_points = PtsUtil.voxel_downsample_point_cloud(points, voxel_size)
-        kdtree = cKDTree(points_normals[:,:3])
-        _, indices = kdtree.query(sampled_points)
-        nearest_points = points_normals[indices]
-
-        normals = nearest_points[:, 3:]
-        camera_axis = -cam_pose[:3, 2] 
-        normals_normalized = normals / np.linalg.norm(normals, axis=1, keepdims=True)
-        cos_theta = np.dot(normals_normalized, camera_axis)
-        theta_rad = np.deg2rad(theta)
-        filtered_sampled_points= sampled_points[cos_theta > np.cos(theta_rad)]
+    def compute_next_best_view_sequence_with_normal(target_point_cloud, target_normal, point_cloud_list, normal_list, scan_points_indices_list, threshold=0.01, overlap_area_threshold=25, init_view = 0, scan_points_threshold=5, status_info=None):
+        selected_views = [init_view]
+        combined_point_cloud = point_cloud_list[init_view]
+        combined_normal = normal_list[init_view]
+        history_indices = [scan_points_indices_list[init_view]]
         
-        return filtered_sampled_points[:, :3]
-    
+        max_rec_pts = np.vstack(point_cloud_list)
+        max_rec_nrm = np.vstack(normal_list)
+        downsampled_max_rec_pts, idx = PtsUtil.voxel_downsample_point_cloud(max_rec_pts, threshold, require_idx=True)
+        downsampled_max_rec_nrm = max_rec_nrm[idx]  
+        max_rec_pts_num = downsampled_max_rec_pts.shape[0]
+        try:
+            max_real_rec_pts_coverage, _ = ReconstructionUtil.compute_coverage_rate_with_normal(target_point_cloud, downsampled_max_rec_pts, target_normal, downsampled_max_rec_nrm, threshold)
+        except:
+            import ipdb; ipdb.set_trace()
+        
+        new_coverage, new_covered_num = ReconstructionUtil.compute_coverage_rate_with_normal(downsampled_max_rec_pts, combined_point_cloud, downsampled_max_rec_nrm, combined_normal, threshold)
+        current_coverage = new_coverage
+        current_covered_num = new_covered_num
+        
+        remaining_views = list(range(len(point_cloud_list)))
+        view_sequence = [(init_view, current_coverage)]
+        cnt_processed_view = 0
+        remaining_views.remove(init_view)
+        curr_rec_pts_num = combined_point_cloud.shape[0]
+        
+        while remaining_views:
+            best_view = None
+            best_coverage_increase = -1
+            best_combined_point_cloud = None
+            best_combined_normal = None
+            best_covered_num = 0
+        
+            for view_index in remaining_views:
+                if point_cloud_list[view_index].shape[0] == 0:
+                    continue
+                if selected_views:
+                    new_scan_points_indices = scan_points_indices_list[view_index]
+                    if not ReconstructionUtil.check_scan_points_overlap(history_indices, new_scan_points_indices, scan_points_threshold):
+                        curr_overlap_area_threshold = overlap_area_threshold
+                    else:
+                        curr_overlap_area_threshold = overlap_area_threshold * 0.5
+                        
+                    if not ReconstructionUtil.check_overlap(point_cloud_list[view_index], combined_point_cloud, overlap_area_threshold = curr_overlap_area_threshold, voxel_size=threshold):
+                        continue
+                
+                new_combined_point_cloud = np.vstack([combined_point_cloud, point_cloud_list[view_index]])
+                new_combined_normal = np.vstack([combined_normal, normal_list[view_index]])
+                new_downsampled_combined_point_cloud, idx = PtsUtil.voxel_downsample_point_cloud(new_combined_point_cloud,threshold, require_idx=True)
+                new_downsampled_combined_normal = new_combined_normal[idx]
+                new_coverage, new_covered_num = ReconstructionUtil.compute_coverage_rate_with_normal(downsampled_max_rec_pts, new_downsampled_combined_point_cloud, downsampled_max_rec_nrm, new_downsampled_combined_normal, threshold)
+                coverage_increase = new_coverage - current_coverage
+                if coverage_increase > best_coverage_increase:
+                    best_coverage_increase = coverage_increase
+                    best_view = view_index
+                    best_covered_num = new_covered_num
+                    best_combined_point_cloud = new_downsampled_combined_point_cloud
+                    best_combined_normal = new_downsampled_combined_normal
+                    
+            
+            if best_view is not None:
+                if best_coverage_increase <=1e-3 or best_covered_num - current_covered_num <= 5:
+                    break
+                
+                selected_views.append(best_view)
+                best_rec_pts_num = best_combined_point_cloud.shape[0]
+                print(f"Current rec pts num: {curr_rec_pts_num}, Best rec pts num: {best_rec_pts_num}, Best cover pts: {best_covered_num}, Max rec pts num: {max_rec_pts_num}")
+                print(f"Current coverage: {current_coverage}, Best coverage increase: {best_coverage_increase}, Max Real coverage: {max_real_rec_pts_coverage}")
+                current_covered_num = best_covered_num
+                curr_rec_pts_num = best_rec_pts_num
+                combined_point_cloud = best_combined_point_cloud
+                combined_normal = best_combined_normal
+                remaining_views.remove(best_view)
+                history_indices.append(scan_points_indices_list[best_view])
+                current_coverage += best_coverage_increase
+                cnt_processed_view += 1
+                if status_info is not None:
+                    sm = status_info["status_manager"]
+                    app_name = status_info["app_name"]
+                    runner_name = status_info["runner_name"]
+                    sm.set_status(app_name, runner_name, "current coverage", current_coverage)
+                    sm.set_progress(app_name, runner_name, "processed view", cnt_processed_view, len(point_cloud_list))
+            
+                view_sequence.append((best_view, current_coverage))
+                
+            else:
+                break
+        if status_info is not None:
+            sm = status_info["status_manager"]
+            app_name = status_info["app_name"]
+            runner_name = status_info["runner_name"]
+            sm.set_progress(app_name, runner_name, "processed view", len(point_cloud_list), len(point_cloud_list))
+        return view_sequence, remaining_views, combined_point_cloud
+
+    
+    @staticmethod
+    def generate_scan_points(display_table_top, display_table_radius, min_distance=0.03, max_points_num = 500, max_attempts = 1000):
+        points = []
+        attempts = 0
+        while len(points) < max_points_num and attempts < max_attempts:
+            angle = np.random.uniform(0, 2 * np.pi)
+            r = np.random.uniform(0, display_table_radius)
+            x = r * np.cos(angle)
+            y = r * np.sin(angle)
+            z = display_table_top
+            new_point = (x, y, z)
+            if all(np.linalg.norm(np.array(new_point) - np.array(existing_point)) >= min_distance for existing_point in points):
+                points.append(new_point)
+            attempts += 1
+        return points
+    
+    @staticmethod
+    def check_scan_points_overlap(history_indices, indices2, threshold=5):
+        for indices1 in history_indices:
+            if len(set(indices1).intersection(set(indices2))) >= threshold:
+                return True
+        return False
+
+        

utils/render.py

@@ -33,12 +33,11 @@ class RenderUtil:
                 print(result.stderr)
                 return None
             path = os.path.join(temp_dir, "tmp")
-            # ------  Debug Start ------
-            import ipdb;ipdb.set_trace()
-            # ------  Debug End ------
             point_cloud = DataLoadUtil.get_target_point_cloud_world_from_path(path, binocular=True)
             cam_params = DataLoadUtil.load_cam_info(path, binocular=True)
-            filtered_point_cloud = ReconstructionUtil.filter_points(point_cloud, model_points_normals, cam_pose=cam_params["cam_to_world"], voxel_size=voxel_threshold, theta=filter_degree)
+            
+            ''' TODO: old code: filter_points api is changed, need to update the code '''
+            filtered_point_cloud = PtsUtil.filter_points(point_cloud, model_points_normals, cam_pose=cam_params["cam_to_world"], voxel_size=voxel_threshold, theta=filter_degree)
             full_scene_point_cloud = None
             if require_full_scene:
                 depth_L, depth_R = DataLoadUtil.load_depth(path, cam_params['near_plane'], cam_params['far_plane'], binocular=True)
@@ -47,7 +46,7 @@ class RenderUtil:
             
                 point_cloud_L = PtsUtil.random_downsample_point_cloud(point_cloud_L, 65536)
                 point_cloud_R = PtsUtil.random_downsample_point_cloud(point_cloud_R, 65536)
-                full_scene_point_cloud = DataLoadUtil.get_overlapping_points(point_cloud_L, point_cloud_R)
+                full_scene_point_cloud = PtsUtil.get_overlapping_points(point_cloud_L, point_cloud_R)
 
             
             return filtered_point_cloud, full_scene_point_cloud

utils/vis.py

@@ -0,0 +1,192 @@
+import numpy as np
+import matplotlib.pyplot as plt
+import sys
+import os
+import trimesh
+
+sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+from utils.data_load import DataLoadUtil
+from utils.pts import PtsUtil
+
+class visualizeUtil:
+    
+    @staticmethod
+    def save_all_cam_pos_and_cam_axis(root, scene, output_dir):
+        length = DataLoadUtil.get_scene_seq_length(root, scene)
+        all_cam_pos = []
+        all_cam_axis = []
+        for i in range(length):
+            path = DataLoadUtil.get_path(root, scene, i)
+            cam_info = DataLoadUtil.load_cam_info(path, binocular=True)
+            cam_pose = cam_info["cam_to_world"]
+            cam_pos = cam_pose[:3, 3]
+            cam_axis = cam_pose[:3, 2] 
+            
+            num_samples = 10
+            sample_points = [cam_pos + 0.02*t * cam_axis for t in range(num_samples)]
+            sample_points = np.array(sample_points)
+        
+            all_cam_pos.append(cam_pos)
+            all_cam_axis.append(sample_points)
+        
+        all_cam_pos = np.array(all_cam_pos)
+        all_cam_axis = np.array(all_cam_axis).reshape(-1, 3)
+        np.savetxt(os.path.join(output_dir, "all_cam_pos.txt"), all_cam_pos)
+        np.savetxt(os.path.join(output_dir, "all_cam_axis.txt"), all_cam_axis)
+        
+    @staticmethod
+    def save_all_combined_pts(root, scene, output_dir):
+        length = DataLoadUtil.get_scene_seq_length(root, scene)
+        all_combined_pts = []   
+        for i in range(length):
+            path = DataLoadUtil.get_path(root, scene, i)
+            pts = DataLoadUtil.load_from_preprocessed_pts(path,"npy")
+            if pts.shape[0] == 0:
+                continue
+            all_combined_pts.append(pts)
+        all_combined_pts = np.vstack(all_combined_pts)
+        downsampled_all_pts = PtsUtil.voxel_downsample_point_cloud(all_combined_pts, 0.001)
+        np.savetxt(os.path.join(output_dir, "all_combined_pts.txt"), downsampled_all_pts)
+
+    @staticmethod
+    def save_seq_cam_pos_and_cam_axis(root, scene, frame_idx_list, output_dir):
+        all_cam_pos = []
+        all_cam_axis = []
+        for i in frame_idx_list:
+            path = DataLoadUtil.get_path(root, scene, i)
+            cam_info = DataLoadUtil.load_cam_info(path, binocular=True)
+            cam_pose = cam_info["cam_to_world"]
+            cam_pos = cam_pose[:3, 3]
+            cam_axis = cam_pose[:3, 2] 
+            
+            num_samples = 10
+            sample_points = [cam_pos + 0.02*t * cam_axis for t in range(num_samples)]
+            sample_points = np.array(sample_points)
+        
+            all_cam_pos.append(cam_pos)
+            all_cam_axis.append(sample_points)
+        
+        all_cam_pos = np.array(all_cam_pos)
+        all_cam_axis = np.array(all_cam_axis).reshape(-1, 3)
+        np.savetxt(os.path.join(output_dir, "seq_cam_pos.txt"), all_cam_pos)
+        np.savetxt(os.path.join(output_dir, "seq_cam_axis.txt"), all_cam_axis)
+        
+    @staticmethod
+    def save_seq_combined_pts(root, scene, frame_idx_list, output_dir):
+        all_combined_pts = []   
+        for i in frame_idx_list:
+            path = DataLoadUtil.get_path(root, scene, i)
+            pts = DataLoadUtil.load_from_preprocessed_pts(path,"npy")
+            if pts.shape[0] == 0:
+                continue
+            all_combined_pts.append(pts)
+        all_combined_pts = np.vstack(all_combined_pts)
+        downsampled_all_pts = PtsUtil.voxel_downsample_point_cloud(all_combined_pts, 0.001)
+        np.savetxt(os.path.join(output_dir, "seq_combined_pts.txt"), downsampled_all_pts)
+        
+    @staticmethod
+    def save_target_mesh_at_world_space(
+        root, model_dir, scene_name, display_table_as_world_space_origin=True
+    ):
+        scene_info = DataLoadUtil.load_scene_info(root, scene_name)
+        target_name = scene_info["target_name"]
+        transformation = scene_info[target_name]
+        if display_table_as_world_space_origin:
+            location = transformation["location"] - DataLoadUtil.get_display_table_top(
+                root, scene_name
+            )
+        else:
+            location = transformation["location"]
+        rotation_euler = transformation["rotation_euler"]
+        pose_mat = trimesh.transformations.euler_matrix(*rotation_euler)
+        pose_mat[:3, 3] = location
+
+        mesh = DataLoadUtil.load_mesh_at(model_dir, target_name, pose_mat)
+        mesh_dir = os.path.join(root, scene_name, "mesh")
+        if not os.path.exists(mesh_dir):
+            os.makedirs(mesh_dir)
+        model_path = os.path.join(mesh_dir, "world_target_mesh.obj")
+        mesh.export(model_path)
+
+    @staticmethod
+    def save_points_and_normals(root, scene, frame_idx, output_dir, binocular=False):
+        target_mask_label = (0, 255, 0, 255)
+        path = DataLoadUtil.get_path(root, scene, frame_idx)
+        cam_info = DataLoadUtil.load_cam_info(path, binocular=binocular, display_table_as_world_space_origin=False)
+        depth = DataLoadUtil.load_depth(
+                path, cam_info["near_plane"], 
+                cam_info["far_plane"], 
+                binocular=binocular,
+            )
+        if isinstance(depth, tuple):
+            depth = depth[0]
+            
+        mask = DataLoadUtil.load_seg(path, binocular=binocular, left_only=True)
+        normal = DataLoadUtil.load_normal(path, binocular=binocular, left_only=True)
+        ''' target points '''
+        if mask is None:
+            target_mask_img = np.ones_like(depth, dtype=bool)
+        else:
+            target_mask_img = (mask == target_mask_label).all(axis=-1)
+        cam_intrinsic = cam_info["cam_intrinsic"]
+        z = depth[target_mask_img]
+        i, j = np.nonzero(target_mask_img)
+        x = (j - cam_intrinsic[0, 2]) * z / cam_intrinsic[0, 0]
+        y = (i - cam_intrinsic[1, 2]) * z / cam_intrinsic[1, 1]
+        
+        random_downsample_N = 1000
+
+        points_camera = np.stack((x, y, z), axis=-1).reshape(-1, 3)
+        normal_camera = normal[target_mask_img].reshape(-1, 3)
+        sampled_target_points, idx = PtsUtil.random_downsample_point_cloud(
+                    points_camera, random_downsample_N, require_idx=True
+                )
+        if len(sampled_target_points) == 0:
+            print("No target points")
+            
+
+        sampled_normal_camera  = normal_camera[idx]
+        sampled_visualized_normal = []
+        sampled_normal_camera[:, 2] = -sampled_normal_camera[:, 2]
+        sampled_normal_camera[:, 1] = -sampled_normal_camera[:, 1]
+        num_samples = 10
+        for i in range(len(sampled_target_points)):
+            sampled_visualized_normal.append([sampled_target_points[i] + 0.02*t * sampled_normal_camera[i] for t in range(num_samples)])
+            
+        sampled_visualized_normal = np.array(sampled_visualized_normal).reshape(-1, 3)
+        np.savetxt(os.path.join(output_dir, "target_pts.txt"), sampled_target_points)
+        np.savetxt(os.path.join(output_dir, "target_normal.txt"), sampled_visualized_normal)
+
+    @staticmethod
+    def save_pts_nrm(root, scene, frame_idx, output_dir, binocular=False):
+        path = DataLoadUtil.get_path(root, scene, frame_idx)
+        pts_world = DataLoadUtil.load_from_preprocessed_pts(path, "npy")
+        nrm_camera = DataLoadUtil.load_from_preprocessed_nrm(path, "npy")
+        cam_info = DataLoadUtil.load_cam_info(path, binocular=binocular)
+        cam_to_world = cam_info["cam_to_world"]
+        nrm_world = nrm_camera @ cam_to_world[:3, :3].T
+        visualized_nrm = []
+        num_samples = 10
+        for i in range(len(pts_world)):
+            for t in range(num_samples):
+                visualized_nrm.append(pts_world[i] - 0.02 * t * nrm_world[i])
+        
+        visualized_nrm = np.array(visualized_nrm)
+        np.savetxt(os.path.join(output_dir, "nrm.txt"), visualized_nrm)
+        np.savetxt(os.path.join(output_dir, "pts.txt"), pts_world)
+        
+# ------ Debug ------
+
+if __name__ == "__main__":
+    root = r"C:\Document\Local Project\nbv_rec\nbv_reconstruction\temp"
+    model_dir = r"H:\\AI\\Datasets\\scaled_object_box_meshes"
+    scene = "box"
+    output_dir = r"C:\Document\Local Project\nbv_rec\nbv_reconstruction\test"
+    
+    #visualizeUtil.save_all_cam_pos_and_cam_axis(root, scene, output_dir)
+    # visualizeUtil.save_all_combined_pts(root, scene, output_dir)
+    # visualizeUtil.save_seq_combined_pts(root, scene, [0, 121, 286, 175, 111,366,45,230,232,225,255,17,199,78,60], output_dir)
+    # visualizeUtil.save_seq_cam_pos_and_cam_axis(root, scene, [0, 121, 286, 175, 111,366,45,230,232,225,255,17,199,78,60], output_dir)
+    # visualizeUtil.save_target_mesh_at_world_space(root, model_dir, scene)
+    #visualizeUtil.save_points_and_normals(root, scene,"10", output_dir, binocular=True)
+    visualizeUtil.save_pts_nrm(root, scene, "116", output_dir, binocular=True)