fix bug: view_sample

data_generator_2.py

@@ -0,0 +1,86 @@
+
+import os
+import bpy
+import numpy as np
+from utils.blender_util import BlenderUtils
+from utils.cad_view_sample_util import CADViewSampleUtil
+from utils.material_util import MaterialUtil
+
+class DataGenerator:
+    def __init__(self, config):
+        self.plane_size = config["runner"]["generate"]["plane_size"] 
+        self.output_dir = config["runner"]["generate"]["output_dir"]
+        self.random_config = config["runner"]["generate"]["random_config"]
+        self.light_and_camera_config = config["runner"]["generate"]["light_and_camera_config"]
+        self.max_views = config["runner"]["generate"]["max_views"]
+        self.min_views = config["runner"]["generate"]["min_views"]
+        self.min_diag = config["runner"]["generate"]["min_diag"]
+        self.max_diag = config["runner"]["generate"]["max_diag"]
+        self.binocular_vision = config["runner"]["generate"]["binocular_vision"]
+        self.target_obj = None
+        self.stopped = False
+        self.random_obj_list = []
+        self.display_table_config = {}
+        BlenderUtils.setup_scene(self.light_and_camera_config, None, self.binocular_vision)
+        self.table = BlenderUtils.get_obj(BlenderUtils.TABLE_NAME)
+        
+    def put_display_object(self, name):
+        obj_mesh_path = BlenderUtils.get_obj_path(self.obj_dir, name)
+        obj = BlenderUtils.load_obj(name, obj_mesh_path)
+        bpy.ops.rigidbody.object_add()
+        bpy.context.object.rigid_body.type = 'ACTIVE'
+        MaterialUtil.change_object_material(obj, MaterialUtil.create_mask_material(color=(0, 1.0, 0)))
+        self.target_obj = obj
+
+    def reset(self):
+        self.target_obj = None
+        self.random_obj_list = []
+        BlenderUtils.reset_objects_and_platform()
+    
+    def start_render(self, diag=0):
+        object_name = self.target_obj.name
+        if "." in object_name:
+            object_name = object_name.split(".")[0]
+        scene_dir = os.path.join(self.output_dir, object_name)
+        if not os.path.exists(scene_dir):
+            os.makedirs(scene_dir)
+        view_num = int(self.min_views + (diag - self.min_diag)/(self.max_diag - self.min_diag) * (self.max_views - self.min_views))
+        view_data = CADViewSampleUtil.sample_view_data_world_space(self.target_obj, distance_range=(0.25,0.5), voxel_size=0.005, max_views=view_num, min_cam_table_included_degree = self.min_cam_table_included_degree, random_view_ratio = self.random_view_ratio )
+        object_points = np.array(view_data["voxel_down_sampled_points"])
+        normals = np.array(view_data["normals"])
+        points_normals = np.concatenate((object_points, normals), axis=1)
+        
+        np.savetxt(os.path.join(scene_dir, "points_and_normals.txt"), points_normals)
+        for i, cam_pose in enumerate(view_data["cam_poses"]):
+            BlenderUtils.set_camera_at(cam_pose)
+            BlenderUtils.render_mask(scene_dir, f"{i}", binocular_vision=self.binocular_vision, target_object = self.target_obj)
+            BlenderUtils.save_cam_params(scene_dir, i, binocular_vision=self.binocular_vision)
+        BlenderUtils.save_scene_info(scene_dir, self.display_table_config, object_name)
+        
+        MaterialUtil.change_object_material(self.target_obj, MaterialUtil.create_normal_material())
+        
+        for i, cam_pose in enumerate(view_data["cam_poses"]):
+            BlenderUtils.set_camera_at(cam_pose)
+            BlenderUtils.render_normal_and_depth(scene_dir, f"{i}", binocular_vision=self.binocular_vision, target_object = self.target_obj)
+            BlenderUtils.save_cam_params(scene_dir, i, binocular_vision=self.binocular_vision)
+        
+        depth_dir = os.path.join(scene_dir, "depth")
+        for depth_file in os.listdir(depth_dir):
+            if not depth_file.endswith(".png"):
+                name, _ = os.path.splitext(depth_file)
+                file_path = os.path.join(depth_dir, depth_file)
+                new_file_path = os.path.join(depth_dir, f"{name}.png")
+                os.rename(file_path,new_file_path)
+        BlenderUtils.save_blend(scene_dir)
+        exit(0)
+        
+        return True
+
+    def gen_scene_data(self, object_name):
+        
+        bpy.context.scene.frame_set(0)
+        self.put_display_object(object_name)
+        diag = BlenderUtils.get_obj_diag(self.target_obj.name)
+        self.start_render(diag)
+    
+        

utils/cad_blender_util.py

@@ -0,0 +1,361 @@
+import os
+import json
+import bpy
+import time
+import gc
+import numpy as np
+import mathutils
+
+
+class CADBlenderUtils:
+
+    TABLE_NAME: str = "table"
+    CAMERA_NAME: str = "Camera"
+    CAMERA_RIGHT_NAME: str = "CameraRight"
+    CAMERA_OBJECT_NAME: str = "CameraObject"
+    DISPLAY_TABLE_NAME: str = "display_table"
+    MESH_FILE_NAME: str = "mesh.obj"
+
+    @staticmethod
+    def get_obj_path(obj_dir, name):
+        return os.path.join(obj_dir, name, CADBlenderUtils.MESH_FILE_NAME)
+
+    @staticmethod
+    def load_obj(name, mesh_path, scale=1):
+        print(mesh_path)
+        bpy.ops.wm.obj_import(filepath=mesh_path)
+        loaded_object = bpy.context.selected_objects[-1]
+        loaded_object.name = name
+        loaded_object.data.name = name
+        loaded_object.scale = (scale, scale, scale)
+        bpy.ops.rigidbody.object_add()
+        return loaded_object
+
+    @staticmethod
+    def get_obj(name):
+        return bpy.data.objects.get(name)
+
+    @staticmethod
+    def get_obj_pose(name):
+        obj = CADBlenderUtils.get_obj(name)
+        return np.asarray(obj.matrix_world)
+
+    @staticmethod
+    def add_plane(name, location, orientation, size=10):
+        bpy.ops.mesh.primitive_plane_add(size=size, location=location)
+        plane = bpy.context.selected_objects[-1]
+        plane.name = name
+        plane.rotation_euler = orientation
+        bpy.ops.rigidbody.object_add()
+        bpy.context.object.rigid_body.type = "PASSIVE"
+
+    @staticmethod
+    def setup_scene(init_light_and_camera_config, table_model_path, binocular_vision):
+        bpy.context.scene.render.engine = "BLENDER_EEVEE"
+        bpy.context.scene.display.shading.show_xray = False
+        bpy.context.scene.display.shading.use_dof = False
+        bpy.context.scene.display.render_aa = "OFF"
+        bpy.context.scene.view_settings.view_transform = "Standard"
+
+        bpy.context.scene.eevee.use_ssr = False  # 关闭屏幕空间反射
+        bpy.context.scene.eevee.use_bloom = False  # 关闭辉光
+        bpy.context.scene.eevee.use_gtao = False  # 关闭环境光遮蔽
+        bpy.context.scene.eevee.use_soft_shadows = False  # 关闭软阴影
+        bpy.context.scene.eevee.use_shadows = False  # 关闭所有阴影
+        bpy.context.scene.world.use_nodes = False  # 如果你不需要环境光，关闭环境节点
+
+        # bpy.context.scene.eevee.use_sss = False            # 关闭次表面散射
+
+        # 2. 设置最低的采样数
+        bpy.context.scene.eevee.taa_render_samples = 1
+        bpy.context.scene.eevee.taa_samples = 1
+        CADBlenderUtils.init_light_and_camera(
+            init_light_and_camera_config, binocular_vision
+        )
+
+    @staticmethod
+    def set_camera_params(camera, config, binocular_vision):
+
+        camera_object = bpy.data.objects.new(CADBlenderUtils.CAMERA_OBJECT_NAME, None)
+        bpy.context.collection.objects.link(camera_object)
+        cameras = [bpy.data.objects.get("Camera")]
+        camera.location = [0, 0, 0]
+        camera.rotation_euler = [0, 0, 0]
+        camera.parent = camera_object
+        if binocular_vision:
+            left_camera = cameras[0]
+            right_camera = left_camera.copy()
+            right_camera.name = CADBlenderUtils.CAMERA_RIGHT_NAME
+            right_camera.data = left_camera.data.copy()
+            right_camera.data.name = CADBlenderUtils.CAMERA_RIGHT_NAME
+            bpy.context.collection.objects.link(right_camera)
+            right_camera.parent = camera_object
+            right_camera.location = [config["eye_distance"] / 2, 0, 0]
+            left_camera.location = [-config["eye_distance"] / 2, 0, 0]
+            binocular_angle = config["eye_angle"]
+            half_angle = np.radians(binocular_angle / 2)
+
+            left_camera.rotation_euler[1] = -half_angle
+            right_camera.rotation_euler[1] = half_angle
+            cameras.append(right_camera)
+
+        for camera in cameras:
+            camera.data.clip_start = config["near_plane"]
+            camera.data.clip_end = config["far_plane"]
+
+            bpy.context.scene.render.resolution_x = config["resolution"][0]
+            bpy.context.scene.render.resolution_y = config["resolution"][1]
+            sensor_height = 24.0
+            focal_length = sensor_height / (
+                2 * np.tan(np.radians(config["fov_vertical"]) / 2)
+            )
+            camera.data.lens = focal_length
+            camera.data.sensor_width = (
+                sensor_height * config["resolution"][0] / config["resolution"][1]
+            )
+            camera.data.sensor_height = sensor_height
+
+    @staticmethod
+    def init_light_and_camera(init_light_and_camera_config, binocular_vision):
+
+        camera = CADBlenderUtils.get_obj(CADBlenderUtils.CAMERA_NAME)
+        CADBlenderUtils.set_camera_params(
+            camera,
+            init_light_and_camera_config[CADBlenderUtils.CAMERA_NAME],
+            binocular_vision,
+        )
+
+    @staticmethod
+    def get_obj_diag(name):
+        obj = CADBlenderUtils.get_obj(name)
+        return np.linalg.norm(obj.dimensions)
+
+    @staticmethod
+    def matrix_to_blender_pose(matrix):
+        location = matrix[:3, 3]
+        rotation_matrix = matrix[:3, :3]
+        rotation_matrix_blender = mathutils.Matrix(rotation_matrix.tolist())
+        rotation_euler = rotation_matrix_blender.to_euler()
+        return location, rotation_euler
+
+    @staticmethod
+    def set_camera_at(pose):
+        camera = CADBlenderUtils.get_obj(CADBlenderUtils.CAMERA_OBJECT_NAME)
+        location, rotation_euler = CADBlenderUtils.matrix_to_blender_pose(pose)
+
+        camera.location = location
+        camera.rotation_euler = rotation_euler
+
+    @staticmethod
+    def get_object_bottom_z(obj):
+        vertices = [v.co for v in obj.data.vertices]
+        vertices_world = [obj.matrix_world @ v for v in vertices]
+        min_z = min([v.z for v in vertices_world])
+        return min_z
+
+    @staticmethod
+    def render_normal_and_depth(
+        output_dir, file_name, binocular_vision=False, target_object=None
+    ):
+        # use pass z
+        bpy.context.scene.view_layers["ViewLayer"].use_pass_z = True
+        target_cameras = [CADBlenderUtils.CAMERA_NAME]
+        if binocular_vision:
+            target_cameras.append(CADBlenderUtils.CAMERA_RIGHT_NAME)
+            
+        for cam_name in target_cameras:
+            bpy.context.scene.camera = CADBlenderUtils.get_obj(cam_name)
+            cam_suffix = "L" if cam_name == CADBlenderUtils.CAMERA_NAME else "R"
+            scene = bpy.context.scene
+            scene.render.filepath = ""
+
+            mask_dir = os.path.join(output_dir, "normal")
+            if not os.path.exists(mask_dir):
+                os.makedirs(mask_dir)
+
+            scene.render.filepath = os.path.join(
+                output_dir, mask_dir, f"{file_name}_{cam_suffix}.exr"
+            ) 
+            
+            scene.render.image_settings.file_format = "OPEN_EXR"
+            scene.render.image_settings.color_mode = "RGB"
+            bpy.context.scene.view_settings.view_transform = "Raw"
+            scene.render.image_settings.color_depth = "16"
+            bpy.context.scene.render.filter_size = 1.5
+            scene.render.resolution_percentage = 100
+            scene.render.use_overwrite = False
+            scene.render.use_file_extension = False
+            scene.render.use_placeholder = False
+            scene.use_nodes = True
+            tree = scene.node_tree
+
+            for node in tree.nodes:
+                tree.nodes.remove(node)
+
+            rl = tree.nodes.new("CompositorNodeRLayers")
+
+            map_range = tree.nodes.new("CompositorNodeMapRange")
+            map_range.inputs["From Min"].default_value = 0.01
+            map_range.inputs["From Max"].default_value = 5
+            map_range.inputs["To Min"].default_value = 0
+            map_range.inputs["To Max"].default_value = 1
+            tree.links.new(rl.outputs["Depth"], map_range.inputs[0])
+
+            output_depth = tree.nodes.new("CompositorNodeOutputFile")
+
+            depth_dir = os.path.join(output_dir, "depth")
+            if not os.path.exists(depth_dir):
+                os.makedirs(depth_dir)
+            output_depth.base_path = depth_dir
+            output_depth.file_slots[0].path = f"{file_name}_{cam_suffix}.####"
+            output_depth.format.file_format = "PNG"
+            output_depth.format.color_mode = "BW"
+            output_depth.format.color_depth = "16"
+            tree.links.new(map_range.outputs[0], output_depth.inputs[0])
+            bpy.ops.render.render(write_still=True)
+
+        msg = "success"
+        return msg
+        
+    @staticmethod
+    def render_mask(
+        output_dir, file_name, binocular_vision=False, target_object=None
+    ):
+        target_cameras = [CADBlenderUtils.CAMERA_NAME]
+        if binocular_vision:
+            target_cameras.append(CADBlenderUtils.CAMERA_RIGHT_NAME)
+        
+        for cam_name in target_cameras:
+            bpy.context.scene.camera = CADBlenderUtils.get_obj(cam_name)
+            cam_suffix = "L" if cam_name == CADBlenderUtils.CAMERA_NAME else "R"
+            scene = bpy.context.scene
+            scene.render.filepath = ""
+
+            mask_dir = os.path.join(output_dir, "mask")
+            if not os.path.exists(mask_dir):
+                os.makedirs(mask_dir)
+
+            scene.render.filepath = os.path.join(
+                output_dir, mask_dir, f"{file_name}_{cam_suffix}.png"
+            )
+            scene.render.image_settings.color_depth = "8"
+            scene.render.resolution_percentage = 100
+            scene.render.use_overwrite = False
+            scene.render.use_file_extension = False
+            scene.render.use_placeholder = False
+
+            
+            bpy.ops.render.render(write_still=True)
+
+        msg = "success"
+        return msg
+
+    @staticmethod
+    def save_cam_params(scene_dir, idx, binocular_vision=False):
+        camera = CADBlenderUtils.get_obj(CADBlenderUtils.CAMERA_NAME)
+        extrinsic = np.array(camera.matrix_world)
+        cam_data = camera.data
+        focal_length = cam_data.lens
+        sensor_width = cam_data.sensor_width
+        sensor_height = cam_data.sensor_height
+        resolution_x = bpy.context.scene.render.resolution_x
+        resolution_y = bpy.context.scene.render.resolution_y
+        intrinsic = np.zeros((3, 3))
+        intrinsic[0, 0] = focal_length * resolution_x / sensor_width  # fx
+        intrinsic[1, 1] = focal_length * resolution_y / sensor_height  # fy
+        intrinsic[0, 2] = resolution_x / 2.0  # cx
+        intrinsic[1, 2] = resolution_y / 2.0  # cy
+        intrinsic[2, 2] = 1.0
+        cam_object = CADBlenderUtils.get_obj(CADBlenderUtils.CAMERA_OBJECT_NAME)
+        extrinsic_cam_object = np.array(cam_object.matrix_world)
+        data = {
+            "extrinsic": extrinsic.tolist(),
+            "extrinsic_cam_object": extrinsic_cam_object.tolist(),
+            "intrinsic": intrinsic.tolist(),
+            "far_plane": camera.data.clip_end,
+            "near_plane": camera.data.clip_start,
+        }
+        if binocular_vision:
+            right_camera = CADBlenderUtils.get_obj(CADBlenderUtils.CAMERA_RIGHT_NAME)
+            extrinsic_right = np.array(right_camera.matrix_world)
+            print("result:", extrinsic_right)
+
+            data["extrinsic_R"] = extrinsic_right.tolist()
+
+        cam_params_dir = os.path.join(scene_dir, "camera_params")
+        if not os.path.exists(cam_params_dir):
+            os.makedirs(cam_params_dir)
+        cam_params_path = os.path.join(cam_params_dir, f"{idx}.json")
+        with open(cam_params_path, "w") as f:
+            json.dump(data, f, indent=4)
+
+    @staticmethod
+    def reset_objects_and_platform():
+        all_objects = bpy.data.objects
+        keep_objects = {
+            "plane_floor",
+            "plane_ceil",
+            "plane_wall_1",
+            "plane_wall_2",
+            "plane_wall_3",
+            "plane_wall_4",
+        }
+        keep_objects.add(CADBlenderUtils.CAMERA_OBJECT_NAME)
+        keep_objects.add(CADBlenderUtils.CAMERA_NAME)
+        keep_objects.add(CADBlenderUtils.CAMERA_RIGHT_NAME)
+        keep_objects.add(CADBlenderUtils.TABLE_NAME)
+
+        for obj in all_objects:
+            if obj.name not in keep_objects:
+                bpy.data.objects.remove(obj, do_unlink=True)
+
+        for block in bpy.data.meshes:
+            if block.users == 0:
+                bpy.data.meshes.remove(block)
+        for block in bpy.data.materials:
+            if block.users == 0:
+                bpy.data.materials.remove(block)
+        for block in bpy.data.images:
+            if block.users == 0:
+                bpy.data.images.remove(block)
+
+        gc.collect()
+        bpy.context.scene.frame_set(0)
+
+    @staticmethod
+    def save_scene_info(scene_root_dir, display_table_config, target_name):
+        all_objects = bpy.data.objects
+        no_save_objects = {
+            "plane_floor",
+            "plane_ceil",
+            "plane_wall_1",
+            "plane_wall_2",
+            "plane_wall_3",
+            "plane_wall_4",
+        }
+        no_save_objects.add(CADBlenderUtils.CAMERA_OBJECT_NAME)
+        no_save_objects.add(CADBlenderUtils.CAMERA_NAME)
+        no_save_objects.add(CADBlenderUtils.CAMERA_RIGHT_NAME)
+        no_save_objects.add(CADBlenderUtils.TABLE_NAME)
+        scene_info = {}
+        for obj in all_objects:
+            if (
+                obj.name not in no_save_objects
+                and obj.name != CADBlenderUtils.DISPLAY_TABLE_NAME
+            ):
+                obj_info = {
+                    "location": list(obj.location),
+                    "rotation_euler": list(obj.rotation_euler),
+                    "scale": list(obj.scale),
+                }
+                scene_info[obj.name] = obj_info
+        scene_info[CADBlenderUtils.DISPLAY_TABLE_NAME] = display_table_config
+        scene_info["target_name"] = target_name
+        scene_info_path = os.path.join(scene_root_dir, "scene_info.json")
+        with open(scene_info_path, "w") as outfile:
+            json.dump(scene_info, outfile)
+            
+    @staticmethod
+    def save_blend(scene_root_dir):
+        blend_path = os.path.join(scene_root_dir, "scene.blend")
+        bpy.ops.wm.save_as_mainfile(filepath=blend_path)

utils/cad_view_sample_util.py

@@ -0,0 +1,146 @@
+
+import numpy as np
+import bmesh
+from collections import defaultdict
+from scipy.spatial.transform import Rotation as R
+from utils.pose import PoseUtil
+from utils.pts import PtsUtil
+import random
+
+class CADViewSampleUtil:
+    @staticmethod
+    def farthest_point_sampling(points, num_samples):
+        num_points = points.shape[0]
+        if num_samples >= num_points:
+            return points, np.arange(num_points)
+        sampled_indices = np.zeros(num_samples, dtype=int)
+        sampled_indices[0] = np.random.randint(num_points)
+        min_distances = np.full(num_points, np.inf)
+        for i in range(1, num_samples):
+            current_point = points[sampled_indices[i - 1]]
+            dist_to_current_point = np.linalg.norm(points - current_point, axis=1)
+            min_distances = np.minimum(min_distances, dist_to_current_point)
+            sampled_indices[i] = np.argmax(min_distances)
+        downsampled_points = points[sampled_indices]
+        return downsampled_points, sampled_indices
+    
+    @staticmethod
+    def voxel_downsample(points, voxel_size):
+        voxel_grid = defaultdict(list)
+        for i, point in enumerate(points):
+            voxel_index = tuple((point // voxel_size).astype(int))
+            voxel_grid[voxel_index].append(i)
+
+        downsampled_points = []
+        downsampled_indices = []
+        for indices in voxel_grid.values():
+            selected_index = indices[0]
+            downsampled_points.append(points[selected_index])
+            downsampled_indices.append(selected_index)
+
+        return np.array(downsampled_points), downsampled_indices
+
+    @staticmethod
+    def sample_view_data(obj, distance_range:tuple = (0.25,0.5), voxel_size:float = 0.005, max_views: int = 1, pertube_repeat:int = 1) -> dict:
+        view_data = {
+            "look_at_points": [],
+            "cam_positions": [],
+        }
+        mesh = obj.data
+        bm = bmesh.new()
+        bm.from_mesh(mesh)
+        bm.verts.ensure_lookup_table()
+        bm.faces.ensure_lookup_table()
+        bm.normal_update()
+        
+        look_at_points = []
+        cam_positions = []
+        normals = []
+        for v in bm.verts:
+            look_at_point = np.array(v.co)
+            
+            view_data["look_at_points"].append(look_at_point)
+            normal = np.zeros(3)
+            for loop in v.link_loops:
+                normal += np.array(loop.calc_normal())
+            normal /= len(v.link_loops)
+            normal = normal / np.linalg.norm(normal)
+            if np.isnan(normal).any():
+                continue
+            if np.dot(normal, look_at_point) < 0:
+                normal = -normal
+            normals.append(normal)
+            
+            for _ in range(pertube_repeat):
+                perturb_angle = np.radians(np.random.uniform(0, 10))
+                perturb_axis = np.random.normal(size=3)
+                perturb_axis /= np.linalg.norm(perturb_axis)
+                rotation_matrix = R.from_rotvec(perturb_angle * perturb_axis).as_matrix()
+                perturbed_normal = np.dot(rotation_matrix, normal)
+                middle_distance = (distance_range[0] + distance_range[1]) / 2
+                perturbed_distance = random.uniform(middle_distance-0.05, middle_distance+0.05)
+                cam_position = look_at_point + perturbed_distance * perturbed_normal
+                look_at_points.append(look_at_point)
+                cam_positions.append(cam_position)
+            
+            
+        bm.free() 
+        look_at_points = np.array(look_at_points)
+        cam_positions = np.array(cam_positions)
+        voxel_downsampled_look_at_points, selected_indices = CADViewSampleUtil.voxel_downsample(look_at_points, voxel_size)
+        voxel_downsampled_cam_positions = cam_positions[selected_indices]
+        voxel_downsampled_normals = np.array(normals)[selected_indices]
+
+        fps_downsampled_look_at_points, selected_indices = CADViewSampleUtil.farthest_point_sampling(voxel_downsampled_look_at_points, max_views*2)
+        fps_downsampled_cam_positions = voxel_downsampled_cam_positions[selected_indices]
+
+        view_data["look_at_points"] = fps_downsampled_look_at_points.tolist()
+        view_data["cam_positions"] = fps_downsampled_cam_positions.tolist()
+        view_data["normals"] = voxel_downsampled_normals
+        view_data["voxel_down_sampled_points"] = voxel_downsampled_look_at_points 
+        return view_data
+    
+    @staticmethod
+    def get_world_points_and_normals(view_data: dict, obj_world_pose: np.ndarray) -> tuple:
+        world_points = []
+        world_normals = []
+        for voxel_down_sampled_points, normal in zip(view_data["voxel_down_sampled_points"], view_data["normals"]):
+            voxel_down_sampled_points_world = obj_world_pose @ np.append(voxel_down_sampled_points, 1.0)
+            normal_world = obj_world_pose[:3, :3] @ normal
+            world_points.append(voxel_down_sampled_points_world[:3])
+            world_normals.append(normal_world)
+        return np.array(world_points), np.array(world_normals)
+    
+    @staticmethod
+    def get_cam_pose(view_data: dict, obj_world_pose: np.ndarray, max_views: int) -> np.ndarray:
+        cam_poses = []
+
+        for look_at_point, cam_position in zip(view_data["look_at_points"], view_data["cam_positions"]):
+            look_at_point_world = obj_world_pose @ np.append(look_at_point, 1.0)
+            cam_position_world = obj_world_pose @ np.append(cam_position, 1.0)
+            look_at_point_world = look_at_point_world[:3]  
+            cam_position_world = cam_position_world[:3]    
+            forward_vector = cam_position_world - look_at_point_world
+            forward_vector /= np.linalg.norm(forward_vector)
+            up_vector = np.array([0, 0, 1])
+            right_vector = np.cross(up_vector, forward_vector)
+            rotation_matrix = np.array([right_vector, up_vector, forward_vector]).T
+            cam_pose = np.eye(4)
+            cam_pose[:3, :3] = rotation_matrix
+            cam_pose[:3, 3] = cam_position_world
+            cam_poses.append(cam_pose)
+        if len(cam_poses) > max_views:
+            cam_points = np.array([cam_pose[:3, 3] for cam_pose in cam_poses])
+            _, indices = PtsUtil.fps_downsample_point_cloud(cam_points, max_views, require_idx=True)
+            cam_poses = [cam_poses[i] for i in indices]
+        
+        return np.array(cam_poses)
+        
+    @staticmethod
+    def sample_view_data_world_space(obj, distance_range:tuple = (0.3,0.5), voxel_size:float = 0.005, max_views: int=1, min_cam_table_included_degree:int=20, random_view_ratio:float = 0.2) -> dict:
+        obj_world_pose = np.asarray(obj.matrix_world)
+        view_data = CADViewSampleUtil.sample_view_data(obj, distance_range, voxel_size, max_views)
+        view_data["cam_poses"] = CADViewSampleUtil.get_cam_pose(view_data, obj_world_pose, max_views, min_cam_table_included_degree, random_view_ratio)
+        view_data["voxel_down_sampled_points"], view_data["normals"] =  CADViewSampleUtil.get_world_points_and_normals(view_data, obj_world_pose)
+        return view_data
+    

utils/view_sample_util.py

@@ -130,14 +130,17 @@ class ViewSampleUtil:
 
             dot_product = np.dot(forward_vector, up_vector)
             angle = np.degrees(np.arccos(dot_product))
+            right_vector = np.cross(up_vector, forward_vector)
             print(angle)
-            if angle < 110:
+            if angle > 90 - min_cam_table_included_degree:
-                target_angle = np.random.uniform(110, 150)
+                max_angle = 90 - min_cam_table_included_degree
+                min_angle = max(90 - min_cam_table_included_degree*2, 30)
+                target_angle = np.random.uniform(min_angle, max_angle)
                 angle_difference = np.radians(target_angle - angle)
 
                 rotation_axis = np.cross(forward_vector, up_vector)
                 rotation_axis /= np.linalg.norm(rotation_axis)
-                rotation_matrix = PoseUtil.rotation_matrix_from_axis_angle(rotation_axis, angle_difference)
+                rotation_matrix = PoseUtil.rotation_matrix_from_axis_angle(rotation_axis, -angle_difference)
                 new_cam_position_world = np.dot(rotation_matrix, cam_position_world - look_at_point_world) + look_at_point_world
                 cam_position_world = new_cam_position_world
                 forward_vector = cam_position_world - look_at_point_world
@@ -148,7 +151,10 @@ class ViewSampleUtil:
                 corrected_up_vector = np.cross(forward_vector, right_vector)
                 rotation_matrix = np.array([right_vector, corrected_up_vector, forward_vector]).T
             else:
-                rotation_matrix = np.array([right_vector, up_vector, forward_vector]).T
+                right_vector = np.cross(up_vector, forward_vector)
+                right_vector /= np.linalg.norm(right_vector)
+                corrected_up_vector = np.cross(forward_vector, right_vector)
+                rotation_matrix = np.array([right_vector, corrected_up_vector, forward_vector]).T
             cam_pose = np.eye(4)
             cam_pose[:3, :3] = rotation_matrix
             cam_pose[:3, 3] = cam_position_world
@@ -162,7 +168,7 @@ class ViewSampleUtil:
                 cos_angle = np.dot(direction_vector, horizontal_normal) / (np.linalg.norm(direction_vector) * np.linalg.norm(horizontal_normal))
                 angle = np.arccos(np.clip(cos_angle, -1.0, 1.0))  
                 angle_degree = np.degrees(angle)
-                if angle_degree < 90 + min_cam_table_included_degree:
+                if angle_degree < 90 - min_cam_table_included_degree:
                     filtered_cam_poses.append(cam_pose)
                 if random.random() < random_view_ratio:
                     pertube_pose = PoseUtil.get_uniform_pose([0.1, 0.1, 0.1], [3, 3, 3], 0, 180, "cm")
@@ -172,7 +178,7 @@ class ViewSampleUtil:
             cam_points = np.array([cam_pose[:3, 3] for cam_pose in filtered_cam_poses])
             _, indices = PtsUtil.fps_downsample_point_cloud(cam_points, max_views, require_idx=True)
             filtered_cam_poses = [filtered_cam_poses[i] for i in indices]
-        
+            
         return np.array(filtered_cam_poses)
         
     @staticmethod