import time import open3d as o3d import os import numpy as np from scipy.spatial.transform import Rotation import sys sys.path.append("/home/algo/Documents/openMVS/openMVS/libs/MVS/utils") from colmap_loader import read_cameras_text, read_images_text, read_int_text, write_int_text, read_indices_from_file from get_pose_matrix import get_w2c import argparse import matplotlib.pyplot as plt import collections import torch import torch.nn.functional as F from torch.utils.dlpack import to_dlpack, from_dlpack import os from typing import Dict, List, Set 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( "--base_path", type=str, required=True, ) parser.add_argument( "--mesh_path", type=str, required=True, ) parser.add_argument( "--sparse_dir", type=str, required=True, ) #""" # print("ModelProcessor Init", args.input_file, self.pose_path) args = parser.parse_args() self.id = args.id self.asset_dir = args.base_path self.mesh_path = args.mesh_path # self.pose_path = f"{self.asset_dir}/sparse/" self.pose_path = args.sparse_dir if not os.path.exists(self.pose_path): raise FileNotFoundError(f"Camera data not found: {self.pose_path}") # GPU设备设置 self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') print(f"Using device: {self.device}") self.mesh = None def load_model(self): """加载并初始化3D模型""" # model_path = f"{self.asset_dir}/repair.ply" model_path = self.mesh_path 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]}") #""" #""" # 将数据转移到GPU vertices = np.asarray(self.mesh.vertices, dtype=np.float32) triangles = np.asarray(self.mesh.triangles, dtype=np.int32) # 转换为PyTorch张量并转移到GPU self.vertices_tensor = torch.from_numpy(vertices).to(self.device) self.triangles_tensor = torch.from_numpy(triangles).to(self.device) print(f"Loaded {len(vertices)} vertices and {len(triangles)} triangles and {len(self.triangles_tensor)} triangles_tensor") self._build_face_adjacency_gpu() #""" # self._build_face_adjacency() 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)) ) def _build_face_adjacency_gpu(self): """优化的GPU版本面片邻接关系构建""" if len(self.triangles_tensor) == 0: return triangles = self.triangles_tensor.cpu().numpy() # 转到CPU处理 num_faces = len(triangles) # 使用更高效的方法构建边-面映射 edge_face_map = {} for face_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(face_idx) # 构建邻接关系 self.face_adjacency = [[] for _ in range(num_faces)] adjacency_count = 0 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: if j not in self.face_adjacency[i]: self.face_adjacency[i].append(j) adjacency_count += 1 print(f"邻接关系构建完成:") print(f"- 面片总数: {num_faces}") print(f"- 边总数: {len(edge_face_map)}") print(f"- 共享边数: {len([f for f in edge_face_map.values() if len(f) > 1])}") print(f"- 邻接关系数: {adjacency_count}") def _build_depth_pyramid_gpu(self, depth_map, levels=4): """GPU版本的深度金字塔构建""" if not isinstance(depth_map, torch.Tensor): depth_tensor = torch.from_numpy(depth_map).float().to(self.device) else: depth_tensor = depth_map.float() pyramid = [depth_tensor] current_level = depth_tensor for _ in range(levels-1): # 使用平均池化进行下采样 current_level = current_level.unsqueeze(0).unsqueeze(0) # 添加batch和channel维度 current_level = F.avg_pool2d(current_level, kernel_size=2, stride=2) current_level = current_level.squeeze(0).squeeze(0) pyramid.append(current_level) return pyramid def _hierarchical_occlusion_test_gpu(self, vertices_cam, depth_pyramid, intrinsics, img_size): """GPU版本的层级遮挡检测 - 直接计算方法""" fx, fy, cx, cy = intrinsics height, width = img_size # 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 (torch.zeros(len(vertices_cam), dtype=torch.bool, device=self.device), torch.zeros(len(vertices_cam), dtype=torch.bool, device=self.device)) visible = torch.zeros(len(vertices_valid), dtype=torch.bool, device=self.device) occlusion = torch.zeros(len(vertices_valid), dtype=torch.bool, device=self.device) # 批量处理所有层级 for level in reversed(range(len(depth_pyramid))): scale = 2 ** level current_depth = depth_pyramid[level] h, w = current_depth.shape # 直接计算投影坐标，避免矩阵乘法 x = vertices_valid[:, 0] y = vertices_valid[:, 1] z = vertices_valid[:, 2] # 缩放的内参 fx_scaled = max(fx/(scale + 1e-6), 1e-6) fy_scaled = max(fy/(scale + 1e-6), 1e-6) cx_scaled = (cx - 0.5)/scale + 0.5 cy_scaled = (cy - 0.5)/scale + 0.5 # 投影计算 u = (x / z) * fx_scaled + cx_scaled v = (y / z) * fy_scaled + cy_scaled # 边界处理 u = torch.clamp(u, 0.0, float(w-1)) v = torch.clamp(v, 0.0, float(h-1)) # 转换为整数索引 u_idx = torch.clamp(torch.floor(u).long(), 0, w-1) v_idx = torch.clamp(torch.floor(v).long(), 0, h-1) # 批量采样深度值 depth_vals = current_depth[v_idx, u_idx] # 批量深度比较 level_tol = 0.0008 * (2 ** level) visible |= (z <= (depth_vals + level_tol)) occlusion |= (z > (depth_vals + level_tol)) # 映射回原始顶点数量 final_visible = torch.zeros(len(vertices_cam), dtype=torch.bool, device=self.device) final_visible[valid_mask] = visible final_occlusion = torch.zeros(len(vertices_cam), dtype=torch.bool, device=self.device) final_occlusion[valid_mask] = occlusion return final_visible, final_occlusion def _compute_vertex_in_frustum_gpu(self, fx, fy, cx, cy, R, eye, height, width, depth_map, qvec, tvec): """GPU版本的视锥体计算和遮挡检测""" print(f"开始 _compute_vertex_in_frustum_gpu") # 直接使用get_w2c，避免重复计算 w2c = get_w2c(qvec, tvec) # 确保w2c是4x4矩阵 if w2c.shape != (4, 4): if w2c.shape == (3, 4): w2c_4x4 = np.eye(4) w2c_4x4[:3, :] = w2c w2c = w2c_4x4 else: raise ValueError(f"w2c matrix has unexpected shape: {w2c.shape}") # 使用GPU张量 vertices = self.vertices_tensor.float() ones = torch.ones(len(vertices), 1, device=self.device) vertices_homo = torch.cat([vertices, ones], dim=1) w2c_tensor = torch.tensor(w2c, device=self.device, dtype=torch.float32) # 简化矩阵乘法 vertices_cam_homo = (w2c_tensor @ vertices_homo.T).T vertices_cam = vertices_cam_homo[:, :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 & (torch.abs(x_ratio) <= tan_fov_x) & (torch.abs(y_ratio) <= tan_fov_y) # 构建深度金字塔 depth_pyramid = self._build_depth_pyramid_gpu(depth_map) # 多级遮挡检测 visible_mask, occlusion_mask = self._hierarchical_occlusion_test_gpu( vertices_cam, depth_pyramid, (fx, fy, cx, cy), (height, width) ) final_visible = torch.zeros(len(vertices), dtype=torch.bool, device=self.device) final_visible[frustum_mask] = visible_mask[frustum_mask] final_occlusion = torch.zeros(len(vertices), dtype=torch.bool, device=self.device) final_occlusion[frustum_mask] = occlusion_mask[frustum_mask] # 转换为numpy数组返回 # return (final_visible.cpu().numpy().tolist(), # self._occlusion_expansion_gpu(final_occlusion, vertices.cpu().numpy())) # 转换为numpy数组返回 return (final_visible.cpu().numpy().tolist()) def _occlusion_expansion_gpu(self, occlusion_mask, vertices, radius=0.0008): """GPU版本的空间哈希遮挡扩展""" if not isinstance(occlusion_mask, torch.Tensor): occlusion_tensor = torch.from_numpy(occlusion_mask).to(self.device) vertices_tensor = torch.from_numpy(vertices).to(self.device) else: occlusion_tensor = occlusion_mask vertices_tensor = torch.from_numpy(vertices).to(self.device) # 构建空间哈希 grid_size = radius * 2 quantized = (vertices_tensor / grid_size).long() # 使用CUDA加速的哈希表 from collections import defaultdict hash_table = defaultdict(list) # 将数据移回CPU进行哈希构建（这部分在GPU上实现较复杂） quantized_cpu = quantized.cpu().numpy() for idx, (x, y, z) in enumerate(quantized_cpu): hash_table[(x, y, z)].append(idx) # 扩展遮挡区域 dilated_mask = occlusion_tensor.cpu().numpy().copy() occluded_indices = np.where(occlusion_tensor.cpu().numpy())[0] for idx in occluded_indices: x, y, z = quantized_cpu[idx] 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) for neighbor_idx in hash_table.get(neighbor_cell, []): dilated_mask[neighbor_idx] = True return dilated_mask.tolist() def _gen_depth_image_gpu(self, cam_data, render): """生成深度图（保持原样，因为Open3D渲染器可能不支持GPU）""" # Open3D的渲染器目前主要在CPU上工作 return self._gen_depth_image(cam_data, render) def _flag_model_gpu(self, camera_data, face_points=None): # 确保使用正确的深度图生成方式 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 - 确保与CPU版本一致 depth_image = self._gen_depth_image_gpu(camera_data, render) # 使用与CPU版本相同的参数计算 R = self.qvec2rotmat(camera_data['qvec']).T eye = -R @ camera_data['tvec'] # final_visible_list, final_occlusion_list = self._compute_vertex_in_frustum_gpu( final_visible_list = self._compute_vertex_in_frustum_gpu( 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'] ) # 确保使用正确的张量设备 final_visible_tensor = torch.tensor(final_visible_list, device=self.device) triangles_tensor = self.triangles_tensor # 直接使用已加载的GPU张量 # 向量化计算面片可见性 v0_indices = triangles_tensor[:, 0] v1_indices = triangles_tensor[:, 1] v2_indices = triangles_tensor[:, 2] v0_visible = final_visible_tensor[v0_indices] v1_visible = final_visible_tensor[v1_indices] v2_visible = final_visible_tensor[v2_indices] face_visible = v0_visible | v1_visible | v2_visible # 使用与CPU版本相同的后续处理 shrunk_visibility = self._shrink_face_visibility(face_visible.cpu().numpy(), 6) expanded_visibility = self._expand_face_visibility(face_visible.cpu().numpy(), 30) shrunk_visibility2 = self._shrink_face_visibility(face_visible.cpu().numpy(), 50) expanded_edge = expanded_visibility & ~shrunk_visibility2 delete_edge = face_visible.cpu().numpy() & ~shrunk_visibility return shrunk_visibility, expanded_edge, delete_edge """ def _gen_depth_image_gpu(self, cam_data, render): # 复制CPU版本的逻辑 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) render.setup_camera(intrinsics, w2c) depth = render.render_to_depth_image(z_in_view_space=True) return np.asarray(depth) # 确保返回numpy数组 """ def _mask_occlusion_gpu(self): """GPU版本的多相机遮挡检测""" cameras = read_cameras_text(os.path.join(self.pose_path, "cameras.txt")) images = read_images_text(os.path.join(self.pose_path, "images.txt")) visible_faces_dict = {} edge_faces_dict = {} delete_edge_faces_dict = {} total_start = time.time() for n, img in enumerate(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(f"处理图像 {img_name} ({n+1}/{len(images)})") # if (img_name!="73_8" and img_name!="52_8" and img_name!="62_8"): # if (img_name!="52_8" and img_name!="62_8"): # if (img_name!="52_8"): # continue start_time = time.time() face_visibility, face_edge, face_delete_edge = self._flag_model_gpu(camera_data) processing_time = time.time() - start_time visible_faces = np.where(face_visibility)[0].tolist() visible_faces_dict[img_name] = visible_faces edge_faces_dict[img_name] = np.where(face_edge)[0].tolist() delete_edge_faces_dict[img_name] = np.where(face_delete_edge)[0].tolist() print(f"图像 {img_name} 处理完成，耗时: {processing_time:.2f}秒，可见面数量{len(visible_faces)}") total_time = time.time() - total_start print(f"所有图像处理完成，总耗时: {total_time:.2f}秒") print(f"平均每张图像耗时: {total_time/len(images):.2f}秒") self.save_occlusion_data(visible_faces_dict, edge_faces_dict, delete_edge_faces_dict, self.asset_dir) return { "result1": visible_faces_dict, "result2": edge_faces_dict, "result3": delete_edge_faces_dict } #""" 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, shrink_radius = 1): 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 = shrink_radius # 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( # visible_mask, occlusion_mask, vertex_depth_difference = self._hierarchical_occlusion_test2( 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 # final_vertex_difference = np.zeros(len(vertices), dtype=bool) # final_vertex_difference[frustum_mask] = vertex_depth_difference # return final_visible.tolist(), self._occlusion_expansion(final_occlusion, vertices), final_vertex_difference.tolist() 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), 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 _hierarchical_occlusion_test2(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), np.zeros(len(vertices_cam), dtype=bool), np.zeros(len(vertices_cam))) # 返回空的深度差值数组 visible = np.zeros(len(vertices_valid), dtype=bool) occlusion = np.zeros(len(vertices_valid), dtype=bool) # 用于存储每个像素点的深度范围（最小值和最大值） pixel_depth_min = {} pixel_depth_max = {} # 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] # 只在最高分辨率层级（level=0）记录像素深度范围 if level == 0: for i in range(len(u_idx)): pixel_key = (u_idx[i], v_idx[i]) vertex_depth = vertices_valid[i, 2] # 更新像素的最小深度值 if pixel_key not in pixel_depth_min or vertex_depth < pixel_depth_min[pixel_key]: pixel_depth_min[pixel_key] = vertex_depth # 更新像素的最大深度值 if pixel_key not in pixel_depth_max or vertex_depth > pixel_depth_max[pixel_key]: pixel_depth_max[pixel_key] = vertex_depth # 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)) # 计算每个像素的深度差值（最大深度 - 最小深度） pixel_depth_difference = {} for pixel_key in pixel_depth_min: if pixel_key in pixel_depth_max: pixel_depth_difference[pixel_key] = pixel_depth_max[pixel_key] - pixel_depth_min[pixel_key] # 为每个顶点分配对应的像素点深度差值 vertex_depth_difference = np.zeros(len(vertices_cam)) if level == 0: # 确保我们记录了深度范围 for i in range(len(vertices_valid)): pixel_key = (u_idx[i], v_idx[i]) if pixel_key in pixel_depth_difference: # 找到原始顶点索引 orig_idx = np.where(valid_mask)[0][i] vertex_depth_difference[orig_idx] = pixel_depth_difference[pixel_key] # 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, vertex_depth_difference def _hierarchical_occlusion_test3(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), np.zeros(len(vertices_cam), dtype=bool), {}) # 返回空的深度差值字典 visible = np.zeros(len(vertices_valid), dtype=bool) occlusion = np.zeros(len(vertices_valid), dtype=bool) # 用于存储每个像素点的深度范围 pixel_depth_range = {} # 用于存储每个顶点对应的像素坐标和深度差值 vertex_pixel_info = {} # 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] # 记录每个像素点的深度范围（只在最高分辨率层级记录） # if level == 0: # 只在原始分辨率层级记录 if True: for i in range(len(u_idx)): vertex_idx = np.where(valid_mask)[0][i] # 获取原始顶点索引 pixel_key = (u_idx[i], v_idx[i]) # 记录顶点对应的像素坐标 vertex_pixel_info[vertex_idx] = pixel_key # 记录像素点的深度范围 if pixel_key not in pixel_depth_range: pixel_depth_range[pixel_key] = { 'min': vertices_valid[i, 2], # 顶点深度 'max': vertices_valid[i, 2], # 顶点深度 'count': 1 } else: pixel_depth_range[pixel_key]['min'] = min( pixel_depth_range[pixel_key]['min'], vertices_valid[i, 2]) pixel_depth_range[pixel_key]['max'] = max( pixel_depth_range[pixel_key]['max'], vertices_valid[i, 2]) pixel_depth_range[pixel_key]['count'] += 1 # 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)) # 计算每个像素点的深度差值 pixel_depth_difference = {} for pixel_key, depth_range in pixel_depth_range.items(): pixel_depth_difference[pixel_key] = depth_range['max'] - depth_range['min'] # 为每个顶点分配对应的像素点深度差值 vertex_depth_difference = np.zeros(len(vertices_cam)) for vertex_idx, pixel_key in vertex_pixel_info.items(): if pixel_key in pixel_depth_difference: vertex_depth_difference[vertex_idx] = pixel_depth_difference[pixel_key] # 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, vertex_depth_difference 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, final_vertex_difference_list = self._compute_vertex_in_frustum( 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)) # 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] ]) shrunk_visibility = self._shrink_face_visibility(face_visible_bitmap, 6) # 6 10 expanded_visibility = self._expand_face_visibility(face_visible_bitmap, 30) shrunk_visibility2 = self._shrink_face_visibility(face_visible_bitmap, 50) expanded_edge = expanded_visibility & ~shrunk_visibility2 delete_edge = face_visible_bitmap & ~shrunk_visibility return shrunk_visibility, expanded_edge, delete_edge def _flag_contour(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) # 获取相机参数 fx = camera_data['fx'] fy = camera_data['fy'] cx = camera_data['cx'] cy = camera_data['cy'] height = camera_data['height'] width = camera_data['width'] # 计算顶点在相机空间中的坐标 w2c = get_w2c(camera_data['qvec'], camera_data['tvec']) vertices = np.asarray(self.mesh.vertices) vertices_homo = np.hstack([vertices, np.ones((len(vertices), 1))]) vertices_cam = (w2c @ vertices_homo.T).T[:, :3] # 过滤掉相机后面的顶点 valid_mask = vertices_cam[:, 2] > 0 vertices_valid = vertices_cam[valid_mask] # 投影顶点到图像平面 u = (vertices_valid[:, 0] * fx / vertices_valid[:, 2] + cx) v = (vertices_valid[:, 1] * fy / vertices_valid[:, 2] + cy) u_idx = np.clip(np.floor(u).astype(int), 0, width-1) v_idx = np.clip(np.floor(v).astype(int), 0, height-1) # 初始化 min_depth_map 和 max_depth_map min_depth_map = np.full((height, width), np.inf) max_depth_map = np.zeros((height, width)) # 更新 min_depth_map 和 max_depth_map for i in range(len(vertices_valid)): x = u_idx[i] y = v_idx[i] d = vertices_valid[i, 2] if d < min_depth_map[y, x]: min_depth_map[y, x] = d if d > max_depth_map[y, x]: max_depth_map[y, x] = d # 对于每个顶点，检查深度范围 edge_vertices = np.zeros(len(vertices), dtype=bool) threshold = 3 # 阈值，可根据需要调整 for i in range(len(vertices_valid)): x = u_idx[i] y = v_idx[i] if min_depth_map[y, x] < np.inf: # 确保有数据 depth_range = max_depth_map[y, x] - min_depth_map[y, x] if depth_range > threshold: # 找到原始顶点索引 orig_idx = np.where(valid_mask)[0][i] edge_vertices[orig_idx] = False # 标记边缘顶点 vertex_colors = np.asarray(self.mesh.vertex_colors) for i in range(len(vertices)): if edge_vertices[i]: vertex_colors[i] = [1.0, 0.0, 0.0] # 红色表示边缘 # 保存模型 output_path = f"{self.asset_dir}/mesh_{self.id}_edge.ply" o3d.io.write_triangle_mesh(output_path, self.mesh) print(f"Edge detection completed. Results saved to {output_path}") # 计算面片的边缘性 triangles = np.asarray(self.mesh.triangles) face_edge = np.zeros(len(triangles), dtype=bool) for face_idx, face in enumerate(triangles): if any(edge_vertices[face]): face_edge[face_idx] = True # 为了兼容原有代码，返回面片可见性和边缘性 # 注意：这里face_visible_bitmap未定义，但原有代码可能期望返回两个值 # 如果需要面片可见性，可以保留原有逻辑，但这里简化处理 face_visible_bitmap = np.ones(len(triangles), dtype=bool) # 临时填充 return face_visible_bitmap, face_edge """ 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 = {} countour_faces_dict = {} visible_faces_dict = {} edge_faces_dict = {} delete_edge_faces_dict = {} total_start = time.time() n = 0 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] # if (img_name!="73_8" and img_name!="52_8" and img_name!="62_8"): # if (img_name!="52_8" and img_name!="62_8"): # if (img_name!="52_8"): # continue start_time = time.time() face_visibility, face_edge, face_delete_edge = self._flag_model(camera_data, None) processing_time = time.time() - start_time visible_faces = np.where(face_visibility)[0].tolist() visible_faces_dict[img.name[:-4]] = visible_faces edge_faces_dict[img.name[:-4]] = np.where(face_edge)[0].tolist() delete_edge_faces_dict[img.name[:-4]] = np.where(face_delete_edge)[0].tolist() n += 1 print(f"图像={img_name}，耗时={processing_time:.2f}秒，可见面数={len(visible_faces)}") total_time = time.time() - total_start print(f"所有图像处理完成，总耗时: {total_time:.2f}秒") print(f"平均每张图像耗时: {total_time/len(images):.2f}秒") self.save_occlusion_data(visible_faces_dict, edge_faces_dict, delete_edge_faces_dict, self.asset_dir) return {"result1": visible_faces_dict, "result2": edge_faces_dict, "result3": delete_edge_faces_dict} def save_occlusion_data(self, result1: Dict[str, List[int]], result2: Dict[str, List[int]], result3: Dict[str, List[int]], base_path: str) -> None: """ 保存遮挡数据到文件 Args: result1: 可见面字典，包含图像名称和对应的可见面列表 result2: 边面字典，包含图像名称和对应的边面列表 result3: 删除边面字典，包含图像名称和对应的删除边面列表 base_path: 基础文件路径 """ print(f"save_occlusion_data {base_path}, {len(result1)}, {len(result2)}, {len(result3)}") # 处理返回的可见面字典 - 转换为图像名到面编号集合的映射 visible_faces_map: Dict[str, Set[int]] = {} for image_name, face_list in result1.items(): visible_faces_map[image_name] = set(face_list) # 计算所有可见面的并集 face_visible_relative: Set[int] = set() for face_set in visible_faces_map.values(): face_visible_relative.update(face_set) # 处理边面字典 edge_faces_map: Dict[str, Set[int]] = {} for image_name, face_list in result2.items(): edge_faces_map[image_name] = set(face_list) # 处理删除边面字典 delete_edge_faces_map: Dict[str, Set[int]] = {} for image_name, face_list in result3.items(): delete_edge_faces_map[image_name] = set(face_list) # 保存 visible_faces_map try: with open(base_path + "_visible_faces_map.txt", "w", encoding='utf-8') as map_file: for image_name, face_set in visible_faces_map.items(): # 写入图像名称和所有面ID，用空格分隔 line = image_name + " " + " ".join(str(face) for face in face_set) + "\n" map_file.write(line) except IOError as e: print(f"Error writing visible_faces_map file: {e}") # 保存 face_visible_relative try: with open(base_path + "_face_visible_relative.txt", "w", encoding='utf-8') as relative_file: for face in face_visible_relative: relative_file.write(str(face) + "\n") except IOError as e: print(f"Error writing face_visible_relative file: {e}") # 保存 edge_faces_map try: with open(base_path + "_edge_faces_map.txt", "w", encoding='utf-8') as map_file2: for image_name, face_set in edge_faces_map.items(): line = image_name + " " + " ".join(str(face) for face in face_set) + "\n" map_file2.write(line) except IOError as e: print(f"Error writing edge_faces_map file: {e}") # 保存 delete_edge_faces_map try: with open(base_path + "_delete_edge_faces_map.txt", "w", encoding='utf-8') as map_file3: for image_name, face_set in delete_edge_faces_map.items(): line = image_name + " " + " ".join(str(face) for face in face_set) + "\n" map_file3.write(line) except IOError as e: print(f"Error writing delete_edge_faces_map file: {e}") def process(self): print("process") self.load_model() try: # 处理物理相机生成遮挡判断 # return self._mask_occlusion() return self._mask_occlusion_gpu() except Exception as e: print(f"Error during processing: {str(e)}") raise if __name__ == "__main__": ModelProcessor().process()