import pandas as pd import os import json import glob import re import math # 1. 基础配置 input_csv = "outputs/phase5a/appendix_B_scene_inventory_submission_ready.csv" output_submission = "outputs/phase5a/appendix_B_scene_inventory_submission_ready_v2.csv" output_audit = "outputs/phase5a/appendix_B_scene_inventory_audit_v2.csv" if not os.path.exists(input_csv): print(f"Error: 找不到基础文件 {input_csv}") exit(1) df = pd.read_csv(input_csv) if 'Resolution' in df.columns: df.rename(columns={'Resolution': 'ResolutionSetting'}, inplace=True) # 尝试探测项目中的评估配置 eval_holdout_rule = None eval_evidence = "unknown_from_files" # 检查配置或脚本中是否存在 --eval 或 holdout config_paths = glob.glob("configs/*.yaml") + glob.glob("scripts/*.py") for cp in config_paths: if not os.path.exists(cp): continue with open(cp, 'r', encoding='utf-8') as f: content = f.read() if '--eval' in content or 'test_every' in content or 'hold' in content: # 3DGS 默认 COLMAP holdout 是 8 eval_holdout_rule = 8 eval_evidence = f"holdout_every_8_inferred_from_project_scripts" break # 2. 辅助函数：统计图片数 def count_images(dir_path): if not os.path.exists(dir_path): return 0 exts = ['*.png', '*.jpg', '*.jpeg', '*.PNG', '*.JPG', '*.JPEG'] count = 0 for ext in exts: count += len(glob.glob(os.path.join(dir_path, ext))) return count def get_colmap_total(base_path): # 尝试按 3DGS 常用规范找 for img_dir in ['images', 'images_2', 'images_4', 'images_8', 'input', 'rgb']: p = os.path.join(base_path, img_dir) c = count_images(p) if c > 0: return c return 0 # 3. 提取引擎 records = [] for index, row in df.iterrows(): s_key = row['SceneKey'] d_name = row['DisplayName'] ds = row['Dataset'] res_set = row['ResolutionSetting'] s_type = row['SceneType'] data_path = "NOT_FOUND" train_views = "NEEDS_SPLIT_RULE" test_views = "NEEDS_SPLIT_RULE" total_views = "UNKNOWN" split_rule = "unknown_from_files" source_evidence = "" notes = "" # 确定物理路径 possible_paths = [] if ds == "NeRF-Synthetic": base = "/root/autodl-tmp/dataset/Synthetic_NeRF_Verified/Synthetic_NeRF" possible_paths = [os.path.join(base, d_name), os.path.join(base, s_key), os.path.join(base, s_key.capitalize())] elif ds == "Mip-NeRF 360": possible_paths = [f"/root/autodl-tmp/dataset/360/{s_key}"] elif ds == "Tanks&Temples": possible_paths = [f"/root/autodl-tmp/dataset/tnt/{s_key}"] elif ds == "Deep Blending": possible_paths = [f"/root/autodl-tmp/dataset/deepblending_clean/{d_name}", f"/root/autodl-tmp/dataset/deepblending_clean/{s_key}"] for p in possible_paths: if os.path.exists(p): data_path = p break if data_path != "NOT_FOUND": # 优先级 A: JSON splits (主要针对 NeRF-Synthetic) train_json = os.path.join(data_path, "transforms_train.json") test_json = os.path.join(data_path, "transforms_test.json") if os.path.exists(train_json) and os.path.exists(test_json): try: with open(train_json, 'r') as f: train_data = json.load(f) train_views = len(train_data['frames']) with open(test_json, 'r') as f: test_data = json.load(f) test_views = len(test_data['frames']) total_views = train_views + test_views # 检查是否存在 val，如果有也加进 total 但不一定加进 test val_json = os.path.join(data_path, "transforms_val.json") if os.path.exists(val_json): with open(val_json, 'r') as f: total_views += len(json.load(f)['frames']) split_rule = "transforms_train/test" source_evidence = f"parsed_from_{train_json}" except Exception as e: notes = f"Error parsing JSON: {e}" else: # 优先级 B/C: 图像目录 / COLMAP 结构 train_dir_count = count_images(os.path.join(data_path, "train")) test_dir_count = count_images(os.path.join(data_path, "test")) if train_dir_count > 0 and test_dir_count > 0: train_views = train_dir_count test_views = test_dir_count total_views = train_views + test_views split_rule = "explicit_train_test_dirs" source_evidence = "parsed_from_image_folders" else: # COLMAP 格式 - 统计 total 然后推断 total = get_colmap_total(data_path) if total > 0: total_views = total # 检查单个 output 目录下是否记录了 metrics metric_file = f"outputs/vanilla_3dgs_{s_key}/metrics_test_iter30000.json" metric_file_bak = f"outputs/vanilla_3dgs_{s_key}_bak/metrics_test_iter30000.json" if not os.path.exists(metric_file) and os.path.exists(metric_file_bak): metric_file = metric_file_bak if eval_holdout_rule: # 3DGS 标准: llffhold=8 means test views are indices % 8 == 0 # 0, 8, 16... tst = len([i for i in range(total) if i % eval_holdout_rule == 0]) trn = total - tst train_views = trn test_views = tst split_rule = eval_evidence source_evidence = "total_images_and_standard_eval_holdout" else: split_rule = "unknown_from_files" source_evidence = "total_images_only_no_split_rule_found" records.append({ "SceneKey": s_key, "DisplayName": d_name, "Dataset": ds, "TrainViews": train_views, "TestViews": test_views, "TotalViews": total_views, "ResolutionSetting": res_set, "SceneType": s_type, "SplitRule": split_rule, "DataPath": data_path, "SourceEvidence": source_evidence, "Notes": notes }) df_res = pd.DataFrame(records) # 4. 导出 os.makedirs("outputs/phase5a", exist_ok=True) sub_cols = ["SceneKey", "DisplayName", "Dataset", "TrainViews", "TestViews", "TotalViews", "ResolutionSetting", "SceneType", "SplitRule"] audit_cols = ["SceneKey", "DisplayName", "Dataset", "TrainViews", "TestViews", "TotalViews", "ResolutionSetting", "SceneType", "SplitRule", "DataPath", "SourceEvidence", "Notes"] df_res[sub_cols].to_csv(output_submission, index=False) df_res[audit_cols].to_csv(output_audit, index=False) print(f"========================================================") print(f"Submission 级表格已保存至: {output_submission}") print(f"Audit 级审计表已保存至: {output_audit}") print(f"========================================================\n") # 5. Markdown 预览 print("=== Appendix B: Scene Inventory (Submission Ready v2) ===") preview_cols = ["SceneKey", "Dataset", "TrainViews", "TestViews", "TotalViews", "ResolutionSetting", "SplitRule"] print(df_res[preview_cols].to_markdown(index=False)) # 6. Sanity Checks print("\n========================================================") print("SANITY CHECKS") print("========================================================") print(f"1. Scene 总数是否正好 31: {len(df_res) == 31} ({len(df_res)})") repro_scenes = pd.read_csv("outputs/phase5a/task_vapre_splatatlas_repro.csv")['scene'].unique().tolist() print(f"2. SceneKey 是否和 repro CSV 完全一致: {set(df_res['SceneKey']) == set(repro_scenes)}") dist = df_res['Dataset'].value_counts().to_dict() print(f"3. Dataset 分布 (预期 12/9/8/2):") for k, v in dist.items(): print(f" - {k}: {v}") nerf_syn = df_res[df_res['Dataset'] == 'NeRF-Synthetic'] syn_check = all(t == 100 for t in nerf_syn['TrainViews']) and all(t == 200 for t in nerf_syn['TestViews']) print(f"4. NeRF-Synthetic (8 scenes) 是否全部为 100 train / 200 test: {syn_check}") resolved_count = len(df_res[df_res['TrainViews'] != 'NEEDS_SPLIT_RULE']) needs_count = len(df_res[df_res['TrainViews'] == 'NEEDS_SPLIT_RULE']) print(f"5. 得到明确 TrainViews/TestViews 的 scene 数量: {resolved_count}") print(f"6. 仍是 NEEDS_SPLIT_RULE 的 scene 数量: {needs_count}") if needs_count > 0: print("7. NEEDS_SPLIT_RULE scene 的 TotalViews:") for _, r in df_res[df_res['TrainViews'] == 'NEEDS_SPLIT_RULE'].iterrows(): print(f" - {r['SceneKey']}: TotalViews = {r['TotalViews']}, Path = {r['DataPath']}") print(f"\n8. 大小写风险探查 (已自动处理的):") for s in ["lego", "drjohnson", "playroom"]: r = df_res[df_res['SceneKey'] == s].iloc[0] print(f" - Key: {s} -> Found Path: {r['DataPath']}") res_check = df_res['ResolutionSetting'].notnull().all() print(f"9. ResolutionSetting 是否全部非空: {res_check}")