import open3d as o3d import os import numpy as np from scipy.spatial.transform import Rotation import sys sys.path.append("/home/algo/Documents/mask_face_occlusion/") from colmap_loader import read_cameras_text, read_images_text, read_int_text, write_int_text, read_indices_from_file from utils.get_pose_matrix import get_w2c import argparse import matplotlib.pyplot as plt import collections class ModelProcessor: def __init__(self): # argv = sys.argv[sys.argv.index("--") + 1:] if "--" in sys.argv else [] parser = argparse.ArgumentParser() parser.add_argument( "--id", required=True, ) # parser.add_argument( # "--mask_image", # type=str, # default="74_8" # 74_8 123_8 75_8 72_8 44_8 104_8 96_8 # ) args = parser.parse_args() self.id = args.id # self.mask_image = args.mask_image self.mesh = None self.asset_dir = f"/home/algo/Documents/openMVS/data/{self.id}" self.pose_path = f"{self.asset_dir}/sparse/" if not os.path.exists(self.pose_path): raise FileNotFoundError(f"Camera data not found: {self.pose_path}") def load_model(self): """加载并初始化3D模型""" model_path = f"{self.asset_dir}/repair.ply" if not os.path.exists(model_path): raise FileNotFoundError(f"Model file not found: {model_path}") print(model_path) mesh_native = o3d.io.read_triangle_mesh(model_path, enable_post_processing=False) # self.mesh = o3d.io.read_triangle_mesh(model_path, enable_post_processing=False) self.mesh = mesh_native #""" print("Open3D去重前顶点数:", len(mesh_native.vertices)) # self.mesh = mesh_native.merge_close_vertices(eps=1e-6) vertices2 = np.asarray(self.mesh.vertices) print("Open3D去重后顶点数:", len(vertices2)) vertices2_sorted = sorted( vertices2.tolist(), key=lambda x: (x[0], x[1], x[2]) ) if not self.mesh.has_vertex_colors(): num_vertices = len(self.mesh.vertices) self.mesh.vertex_colors = o3d.utility.Vector3dVector( np.ones((num_vertices, 3)) ) self.uv_array = np.asarray(self.mesh.triangle_uvs) # print(f"UV 坐标形状：{self.uv_array.shape}, {self.uv_array[0][1]}") #""" self._build_face_adjacency() #""" def _build_face_adjacency(self): if not self.mesh.triangles: return triangles = np.asarray(self.mesh.triangles) num_faces = len(triangles) self.face_adjacency = [[] for _ in range(num_faces)] # 创建边到面片的映射 edge_face_map = {} for idx, tri in enumerate(triangles): # 获取三条边（排序顶点保证唯一性） edges = [ tuple(sorted([tri[0], tri[1]])), tuple(sorted([tri[1], tri[2]])), tuple(sorted([tri[2], tri[0]])) ] for edge in edges: if edge not in edge_face_map: edge_face_map[edge] = [] edge_face_map[edge].append(idx) # 通过共享边建立邻接关系 for edge, faces in edge_face_map.items(): if len(faces) > 1: # 只处理共享边 for i in faces: for j in faces: if i != j and j not in self.face_adjacency[i]: self.face_adjacency[i].append(j) def _expand_face_visibility(self, face_visibility): if self.face_adjacency is None: return face_visibility.copy() # 使用队列实现广度优先搜索的多层扩展 expanded = face_visibility.copy() visited = set() queue = collections.deque() # 初始添加所有可见面片 for face_idx, is_visible in enumerate(face_visibility): if is_visible: queue.append((face_idx, 0)) # (面片索引, 当前扩展层数) visited.add(face_idx) self.expand_radius = 10 # 广度优先扩展 while queue: current_idx, current_radius = queue.popleft() # 如果当前扩展层数小于目标半径，继续扩展 if current_radius < self.expand_radius: for neighbor_idx in self.face_adjacency[current_idx]: # 仅处理未访问过的面片 if neighbor_idx not in visited: expanded[neighbor_idx] = True visited.add(neighbor_idx) # 将邻居加入队列，扩展层数+1 queue.append((neighbor_idx, current_radius + 1)) return expanded def _shrink_face_visibility(self, face_visibility, shrink_radius=1): if self.face_adjacency is None or shrink_radius == 0: return face_visibility.copy() # 创建当前可见性副本 current_visible = face_visibility.copy() # 创建边界队列 boundary_queue = collections.deque() # 初始化：找出所有边界面片（可见但至少有一个邻居不可见） for idx, is_visible in enumerate(current_visible): if not is_visible: continue for neighbor_idx in self.face_adjacency[idx]: if not current_visible[neighbor_idx]: boundary_queue.append((idx, 1)) # (面片索引, 当前圈数) break # 分层剥离 removed = set() while boundary_queue: idx, current_radius = boundary_queue.popleft() # 如果当前面片已被移除，跳过 if idx in removed: continue # 如果当前圈数已达到目标圈数，标记为移除 if current_radius <= shrink_radius: current_visible[idx] = False removed.add(idx) # 检查邻居：如果邻居是可见的，且尚未被标记为边界 for neighbor_idx in self.face_adjacency[idx]: if current_visible[neighbor_idx] and neighbor_idx not in removed: # 如果邻居现在成为边界（因为当前面片被移除） is_boundary = False for n_neighbor_idx in self.face_adjacency[neighbor_idx]: if not current_visible[n_neighbor_idx]: is_boundary = True break if is_boundary: boundary_queue.append((neighbor_idx, current_radius + 1)) return current_visible #""" @staticmethod def qvec2rotmat(qvec): """四元数转旋转矩阵""" return Rotation.from_quat([qvec[1], qvec[2], qvec[3], qvec[0]]).as_matrix() def _compute_vertex_in_frustum(self, fx, fy, cx, cy, R, eye, height, width, depth_map, qvec, tvec): """基于深度金字塔的层级式遮挡检测""" # 坐标转换 R = self.qvec2rotmat(qvec) w2c = get_w2c(qvec, tvec) vertices = np.asarray(self.mesh.vertices, dtype=np.float32) vertices_homo = np.hstack([vertices, np.ones((len(vertices), 1))]) vertices_cam = (w2c @ vertices_homo.T).T[:, :3] # 视锥体快速剔除 valid_z = vertices_cam[:, 2] > 0 tan_fov_x = (width / 2) / fx tan_fov_y = (height / 2) / fy x_ratio = vertices_cam[:, 0] / vertices_cam[:, 2] y_ratio = vertices_cam[:, 1] / vertices_cam[:, 2] frustum_mask = valid_z & (np.abs(x_ratio) <= tan_fov_x) & (np.abs(y_ratio) <= tan_fov_y) # 构建深度金字塔 depth_pyramid = self._build_depth_pyramid(depth_map) # 多级遮挡检测 visible_mask, occlusion_mask = self._hierarchical_occlusion_test( vertices_cam[frustum_mask], depth_pyramid, (fx, fy, cx, cy), (height, width) ) final_visible= np.zeros(len(vertices), dtype=bool) final_visible[frustum_mask] = visible_mask final_occlusion = np.zeros(len(vertices), dtype=bool) final_occlusion[frustum_mask] = occlusion_mask return final_visible.tolist(), self._occlusion_expansion(final_occlusion, vertices) def _build_depth_pyramid2(self, depth_map, levels=4): """构建深度图金字塔""" pyramid = [depth_map.copy()] current_level = depth_map for _ in range(levels-1): current_level = 0.25 * (current_level[::2, ::2] + current_level[1::2, ::2] + current_level[::2, 1::2] + current_level[1::2, 1::2]) pyramid.append(current_level) return pyramid def _build_depth_pyramid(self, depth_map, levels=4): pyramid = [depth_map.copy()] current_level = depth_map for _ in range(levels-1): h, w = current_level.shape # 确保尺寸可被2整除 if h % 2 != 0 or w % 2 != 0: current_level = current_level[:h//2 * 2, :w//2 * 2] # 裁剪到最近偶尺寸 # 添加广播兼容性检查 if current_level[::2, ::2].shape != current_level[1::2, ::2].shape: current_level = current_level[:h//2 * 2, :w//2 * 2] current_level = 0.25 * ( current_level[::2, ::2] + current_level[1::2, ::2] + current_level[::2, 1::2] + current_level[1::2, 1::2] ) pyramid.append(current_level) return pyramid def _hierarchical_occlusion_test(self, vertices_cam, depth_pyramid, intrinsics, img_size): """层级式遮挡检测（安全版本）""" fx, fy, cx, cy = intrinsics height, width = img_size # 1. 过滤无效顶点 valid_mask = vertices_cam[:, 2] > 1e-6 vertices_valid = vertices_cam[valid_mask] if len(vertices_valid) == 0: return np.zeros(len(vertices_cam), dtype=bool) visible = np.zeros(len(vertices_valid), dtype=bool) occlusion = np.zeros(len(vertices_valid), dtype=bool) # 2. 层级检测 for level in reversed(range(len(depth_pyramid))): scale = 2 ** level current_depth = depth_pyramid[level] h, w = current_depth.shape # 安全构造内参矩阵 K = np.array([ [max(fx/(scale + 1e-6), 1e-6), 0, (cx - 0.5)/scale + 0.5], [0, max(fy/(scale + 1e-6), 1e-6), (cy - 0.5)/scale + 0.5], [0, 0, 1] ], dtype=np.float32) # 投影计算 uv_homo = (K @ vertices_valid.T).T uv = uv_homo[:, :2] / uv_homo[:, 2:3] # 安全边界处理 u = np.clip(uv[:, 0], 0.0, float(w-1)) v = np.clip(uv[:, 1], 0.0, float(h-1)) # 转换为整数索引 u_idx = np.clip(np.floor(u).astype(np.int32), 0, w-1) v_idx = np.clip(np.floor(v).astype(np.int32), 0, h-1) # 采样深度值 depth_vals = current_depth[v_idx, u_idx] # 深度比较 level_tol = 0.0008 * (2 ** level) # 0.005 0.0008 visible |= (vertices_valid[:, 2] <= (depth_vals + level_tol)) occlusion |= (vertices_valid[:, 2] > (depth_vals + level_tol)) # 3. 结果映射 final_visible = np.zeros(len(vertices_cam), dtype=bool) final_visible[valid_mask] = visible final_occlusion = np.zeros(len(vertices_cam), dtype=bool) final_occlusion[valid_mask] = occlusion return final_visible, final_occlusion def _occlusion_expansion(self, occlusion_mask, vertices, radius=0.0008): """基于空间哈希的快速遮挡扩展""" from collections import defaultdict # 构建空间哈希 grid_size = radius * 2 hash_table = defaultdict(list) # 量化顶点坐标 quantized = (vertices / grid_size).astype(int) for idx, (x, y, z) in enumerate(quantized): hash_table[(x, y, z)].append(idx) # 扩展遮挡区域 dilated_mask = occlusion_mask.copy() occluded_indices = np.where(occlusion_mask)[0] for idx in occluded_indices: x, y, z = quantized[idx] # 查询邻近27个网格 for dx in (-1, 0, 1): for dy in (-1, 0, 1): for dz in (-1, 0, 1): neighbor_cell = (x+dx, y+dy, z+dz) dilated_mask[hash_table.get(neighbor_cell, [])] = True return dilated_mask.tolist() def _gen_depth_image(self, cam_data, render): """生成深度图""" qvec = cam_data['qvec'] tvec = cam_data['tvec'] fx = cam_data['fx'] fy = cam_data['fy'] cx = cam_data['cx'] cy = cam_data['cy'] width = cam_data['width'] height = cam_data['height'] intrinsics = o3d.camera.PinholeCameraIntrinsic( width, height, fx=fx, fy=fy, cx=cx, cy=cy) w2c = get_w2c(qvec, tvec) # print(np.linalg.inv(w2c)) # 配置渲染器 render.setup_camera(intrinsics, w2c) depth = render.render_to_depth_image(z_in_view_space=True) return np.asarray(depth) def sort_vertices(vertices_original): return sorted( (v for v in vertices_original), key=lambda v: (v.co.x, v.co.y, v.co.z) ) def _flag_model(self, camera_data, face_points): """标记可见顶点""" vertex_visible = [] vertex_occlusion = [] depth_images = [] render = o3d.visualization.rendering.OffscreenRenderer(camera_data['width'], camera_data['height']) material = o3d.visualization.rendering.MaterialRecord() render.scene.add_geometry("mesh", self.mesh, material) # 生成深度图 depth_image = self._gen_depth_image(camera_data, render) # 计算可见性 R = self.qvec2rotmat(camera_data['qvec']).T eye = -R @ camera_data['tvec'] # eye = camera_data['tvec'] final_visible_list, final_occlusion_list = self._compute_vertex_in_frustum( camera_data['fx'], camera_data['fy'], camera_data['cx'], camera_data['cy'], R, eye, camera_data['height'], camera_data['width'], depth_image,camera_data['qvec'], camera_data['tvec'] ) print("_flag_model", len(final_occlusion_list), len(self.mesh.vertices), len(self.mesh.vertex_colors)) """ vertices = np.asarray(self.mesh.vertices) vertex_index_map = {tuple(v.tolist()): i for i, v in enumerate(vertices)} vertex_colors = np.asarray(self.mesh.vertex_colors) if face_points==None: for vertex_id, coord in enumerate(self.mesh.vertices): if final_occlusion_list[vertex_id]: vertex_colors[vertex_id] = [1.0, 0.0, 0.0] # 保存最终模型 output_path = f"{self.asset_dir}/mesh_{self.id}_遮挡判断.ply" o3d.io.write_triangle_mesh(output_path, self.mesh) print(f"Processing completed. Results saved to {output_path}") else: list_id = [] # sorted_verts = self.sort_vertices(self.mesh.vertices) sorted_verts =sorted( (tuple(v.tolist()) for v in vertices), key=lambda v: (v[0], v[1], v[2]) ) dict_s_o = {} dict_o_s = {} for sorted_idx, sorted_v in enumerate(sorted_verts): original_idx = vertex_index_map[sorted_v] dict_s_o[sorted_idx] = original_idx dict_o_s[original_idx] = sorted_idx for vertex_id, coord in enumerate(self.mesh.vertices): # print(vertex_id, coord) if final_occlusion_list[vertex_id]: if dict_o_s[vertex_id] in face_points: list_id.append(dict_o_s[vertex_id]) vertex_colors[vertex_id] = [1.0, 0.0, 0.0] write_int_text(list_id, os.path.join(self.pose_path, "face_points_mask.txt")) # 保存最终模型 output_path = f"{self.asset_dir}/mesh_{self.id}_脸部遮挡判断.ply" o3d.io.write_triangle_mesh(output_path, self.mesh) print(f"Processing completed. Results saved to {output_path}") """ # 获取三角形面片数组 triangles = np.asarray(self.mesh.triangles) face_visible_bitmap = np.zeros(len(triangles), dtype=bool) # 遍历所有面片 for face_idx, face in enumerate(triangles): v0, v1, v2 = face face_visible_bitmap[face_idx] = any([ # any all final_visible_list[v0], final_visible_list[v1], final_visible_list[v2] ]) # return face_visible_bitmap # expanded_visibility = self._expand_face_visibility(face_visible_bitmap) # return expanded_visibility shrunk_visibility = self._shrink_face_visibility(face_visible_bitmap, 6) # 6 10 return shrunk_visibility """ def _mask_face_occlusion(self): # 读取相机数据 cameras = read_cameras_text(os.path.join(self.pose_path, "cameras.txt")) images = read_images_text(os.path.join(self.pose_path, "images.txt")) # cameras = read_cameras_text(os.path.join(self.pose_path, "backup_cameras.txt")) # images = read_images_text(os.path.join(self.pose_path, "backup_images.txt")) face_points_sorted_path = os.path.join(self.pose_path, "face_points_sorted.txt") print("face_points_sorted_path=", face_points_sorted_path) #face_points = read_int_text(face_points_sorted_path) face_points = read_indices_from_file(face_points_sorted_path) # face_points = {} camera_data = {} for img in images.values(): if self.mask_image == img.name[:-4]: camera = cameras[img.camera_id] camera_data = { "qvec": img.qvec, "tvec": img.tvec, "fx": camera.params[0], "fy": camera.params[1], "cx": camera.params[2], "cy": camera.params[3], "width": camera.width, "height": camera.height, "name": img.name[:-4] } # print(face_points) self._flag_model(camera_data, face_points) """ def _mask_occlusion(self): # 读取相机数据 cameras = read_cameras_text(os.path.join(self.pose_path, "cameras.txt")) images = read_images_text(os.path.join(self.pose_path, "images.txt")) camera_data = {} """ for img in images.values(): if self.mask_image == img.name[:-4]: camera = cameras[img.camera_id] camera_data = { "qvec": img.qvec, "tvec": img.tvec, "fx": camera.params[0], "fy": camera.params[1], "cx": camera.params[2], "cy": camera.params[3], "width": camera.width, "height": camera.height, "name": img.name[:-4] } return self._flag_model(camera_data, None) """ visible_faces_dict = {} for img in images.values(): camera = cameras[img.camera_id] camera_data = { "qvec": img.qvec, "tvec": img.tvec, "fx": camera.params[0], "fy": camera.params[1], "cx": camera.params[2], "cy": camera.params[3], "width": camera.width, "height": camera.height, "name": img.name[:-4] } img_name = img.name[:-4] print("img_name=", img_name) # if (img_name!="72_2" and img_name!="82_2" and img_name!="83_2"): # 82_2 72_2 #if (img_name!="82_2"): # continue face_visibility = self._flag_model(camera_data, None) visible_faces_dict[img.name[:-4]] = np.where(face_visibility)[0].tolist() return visible_faces_dict def process(self): print("process") self.load_model() try: # 处理物理相机生成遮挡判断 # self._mask_face_occlusion() return self._mask_occlusion() except Exception as e: print(f"Error during processing: {str(e)}") raise if __name__ == "__main__": ModelProcessor().process()