| import os |
| import json |
|
|
| class NpEncoder(json.JSONEncoder): |
| def default(self, obj): |
| if hasattr(obj, 'item'): |
| return obj.item() |
| return super().default(obj) |
|
|
| import argparse |
| import shutil |
| import numpy as np |
| import torch |
| import sys |
|
|
| sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..'))) |
| from core.registry import METHOD_REGISTRY |
| import methods |
| from photometric import compute_photometric_metrics |
| from geometric import compute_geometric_pathology |
| from alignment import assess_manifold_collapse |
|
|
| def load_cameras_from_json(cam_path): |
| if not os.path.exists(cam_path): return [] |
| with open(cam_path, 'r') as f: cams_data = json.load(f) |
| class DummyCam: |
| def __init__(self, center): |
| self.camera_center = torch.tensor(center, dtype=torch.float32) |
| return [DummyCam(c['center']) for c in cams_data if 'center' in c] |
|
|
| def parse_args(): |
| parser = argparse.ArgumentParser() |
| parser.add_argument("--method", type=str, required=True) |
| parser.add_argument("--scene", type=str, required=True) |
| parser.add_argument("--render_dir", type=str, required=True) |
| parser.add_argument("--gt_dir", type=str, required=True) |
| parser.add_argument("--ply_path", type=str, required=True) |
| parser.add_argument("--output_json", type=str, required=True) |
| parser.add_argument("--prior_type", type=str, default="none") |
| parser.add_argument("--depth_dir", type=str, default="") |
| parser.add_argument("--skip_quarantine", action="store_true") |
| parser.add_argument("--colmap_dir", type=str, default="") |
| return parser.parse_args() |
|
|
| def main(): |
| args = parse_args() |
| os.makedirs(os.path.dirname(args.output_json), exist_ok=True) |
| |
| |
| model_path = args.ply_path.split('/point_cloud/')[0] |
| try: |
| iteration = int(args.ply_path.split('/iteration_')[1].split('/')[0]) |
| except IndexError: |
| iteration = 30000 |
|
|
| |
| model_class = methods.load_method(args.method) |
| _OUTDOOR_360 = {"bicycle", "flowers", "garden", "stump", "treehill"} |
| _res = 4 if args.scene in _OUTDOOR_360 else 2 |
| dataset_config = {"source_path": args.colmap_dir, "model_path": model_path, "resolution": _res} |
| model = model_class(dataset_config, hyperparams={}) |
| |
| try: |
| model.load(model_path, iteration) |
| except Exception as e: |
| print(f"❌ [Eval] 模型资产加载失败: {e}") |
| return |
|
|
| cams = load_cameras_from_json(os.path.join(model_path, "cameras.json")) |
| |
| |
| geo_metrics = compute_geometric_pathology(model, cameras=cams) |
| flags = geo_metrics.get("pathology_flags", []) |
| |
| |
| colmap_ply_path = os.path.join(args.colmap_dir, "sparse", "0", "points3D.ply") if args.colmap_dir else None |
| alignment_res = assess_manifold_collapse(model, model_path, colmap_ply_path=colmap_ply_path) |
| |
| if alignment_res.get("manifold_collapse", False): |
| flags.append("MANIFOLD_COLLAPSE") |
| pass |
| else: |
| del model; torch.cuda.empty_cache() |
| photo_metrics = compute_photometric_metrics(args.render_dir, args.gt_dir) |
| |
| if args.depth_dir and os.path.exists(args.depth_dir): |
| depth_vars = [np.var(np.load(os.path.join(args.depth_dir, f))[np.load(os.path.join(args.depth_dir, f)) > 0.05]) |
| for f in os.listdir(args.depth_dir) if f.endswith('.npy')] |
| mean_depth_var = np.mean(depth_vars) if depth_vars else 1.0 |
| geo_metrics["mean_depth_variance"] = round(float(mean_depth_var), 4) |
| if mean_depth_var < 0.01: |
| flags.append("FLAT_WALL_COLLAPSE") |
| |
| geo_metrics["pathology_flags"] = flags |
| |
| report = { |
| "method": args.method, |
| "scene": args.scene, |
| "prior_type": args.prior_type, |
| "photometric": photo_metrics, |
| "geometric": geo_metrics, |
| "pathology_flags": list(set(flags)) |
| } |
| |
| with open(args.output_json, 'w') as f: |
| json.dump(report, f, indent=4) |
| |
| if len(flags) > 0 and not args.skip_quarantine: |
| master_log_path = os.path.join(os.path.dirname(model_path), "pathology_master_log.txt") |
| with open(master_log_path, "a") as log_f: |
| log_f.write(f"Method: {args.method} | Scene: {args.scene} | Path: {model_path} | Flags: {','.join(list(set(flags)))}\n") |
|
|
| if __name__ == "__main__": |
| main() |
|
|
