#!/usr/bin/env python3 """ 核查 benchmark 标注内的物体实例是否与原标注文件（corrected）对应。 Metadata 规范来源（用户指定）： - Part1 长视频：https://huggingface.co/datasets/Torwnexial/ready4label/tree/main/new_long_video/corrected_json_2 - Part2+3 中短视频：https://huggingface.co/datasets/Torwnexial/ready4label/tree/main/top20merge_full/corrected_json_2 本地比对请使用从上述 HF 同步的目录（如 data/hf_ready4label/.../corrected_json_2）。 原标注格式：corrected_json_2 下每视频一个 JSON，含 frames[]，每帧含 frame_idx, clip_idx, objects[]， 每个 object 含 concept。 用法: python verify_annotation_against_source.py \\ --bench-dir benchmark_single \\ --raw-dir data/hf_ready4label/new_long_video/corrected_json_2 \\ --part 1 """ from __future__ import annotations import argparse import json import os import sys from pathlib import Path from typing import Any, Dict, List, Optional, Tuple def load_json(path: str) -> Dict[str, Any]: with open(path, encoding="utf-8") as f: return json.load(f) def raw_per_frame_objects(raw: Dict[str, Any]) -> Dict[int, List[str]]: """原标注: frame_idx -> 该帧内物体 concept 列表（按出现顺序，可重复）。""" out: Dict[int, List[str]] = {} for fr in raw.get("frames") or []: fidx = fr.get("frame_idx") if fidx is None: continue objs = fr.get("objects") or [] out[fidx] = [o.get("concept", "").strip() for o in objs if isinstance(o, dict) and o.get("concept")] return out def raw_concept_count_up_to_clip(raw: Dict[str, Any], clip_idx: int) -> Dict[str, int]: """原标注: 到 clip_idx（含）为止，每个 concept 的物体实例总数（每帧内 objects 条数累加）。""" counts: Dict[str, int] = {} for fr in raw.get("frames") or []: if fr.get("clip_idx", 0) > clip_idx: continue for o in fr.get("objects") or []: if not isinstance(o, dict): continue c = (o.get("concept") or "").strip() if c: counts[c] = counts.get(c, 0) + 1 return counts def compare_counting( bench_video: Dict[str, Any], raw: Dict[str, Any], video_name: str, ) -> List[Tuple[str, str, Any, Any, str]]: """返回 [(video_name, concept, bench_answer, raw_count, msg), ...] 差异列表。""" diffs: List[Tuple[str, str, Any, Any, str]] = [] for t in bench_video.get("tasks") or []: if t.get("task_type") != "object_counting": continue concept = (t.get("concept") or "").strip() if not concept: continue cps = t.get("checkpoints") or [] if not cps: continue cp = cps[0] clip_idx = cp.get("clip_idx", 0) bench_answer = cp.get("answer") raw_counts = raw_concept_count_up_to_clip(raw, clip_idx) raw_count = raw_counts.get(concept, 0) if bench_answer is not None and raw_count != bench_answer: diffs.append((video_name, concept, bench_answer, raw_count, "counting")) return diffs def normalize_objects_list(lst: List[str]) -> List[str]: return sorted(lst) # 比较时忽略顺序 def compare_frame_recall( bench_video: Dict[str, Any], raw_by_frame: Dict[int, List[str]], video_name: str, ) -> List[Tuple[str, int, List[str], List[str], str]]: """返回 [(video_name, frame_idx, bench_objs, raw_objs, msg), ...] 差异列表。""" diffs: List[Tuple[str, int, List[str], List[str], str]] = [] for t in bench_video.get("tasks") or []: if t.get("task_type") != "frame_recall": continue cps = t.get("checkpoints") or [] if not cps: continue cp = cps[0] for f in cp.get("frames") or []: fidx = f.get("frame_idx") if fidx is None: continue bench_objs = f.get("objects_in_frame") or [] raw_objs = raw_by_frame.get(fidx, []) if normalize_objects_list(bench_objs) != normalize_objects_list(raw_objs): diffs.append((video_name, fidx, bench_objs, raw_objs, "frame_recall")) return diffs def main() -> None: ap = argparse.ArgumentParser(description="核查 benchmark 与原标注物体实例是否对应") ap.add_argument("--bench-dir", type=str, required=True, help="benchmark_single 目录") ap.add_argument("--raw-dir", type=str, required=True, help="原标注 corrected_json_2 目录（每视频一个 JSON）") ap.add_argument("--part", type=str, choices=["1", "2", "3"], default="1", help="Part 1/2/3") ap.add_argument("--max-videos", type=int, default=None, help="最多核查视频数，默认全部") args = ap.parse_args() if args.part == "1": bench_file = os.path.join(args.bench_dir, "part1_long_videos_-_dual_format_appearance.json") elif args.part == "2": bench_file = os.path.join(args.bench_dir, "part2_short_videos_-_place_&_motion.json") else: bench_file = os.path.join(args.bench_dir, "part3_short_videos_-_objects_with_dual_format_fixed_choices.json") if not os.path.isfile(bench_file): print(f"Error: benchmark not found: {bench_file}", file=sys.stderr) sys.exit(1) if not os.path.isdir(args.raw_dir): print(f"Error: raw dir not found: {args.raw_dir}", file=sys.stderr) sys.exit(1) data = load_json(bench_file) videos = data.get("videos") or data.get("data") or [] max_v = args.max_videos or len(videos) all_count_diffs: List[Tuple[str, str, Any, Any, str]] = [] all_fr_diffs: List[Tuple[str, int, List[str], List[str], str]] = [] checked = 0 missing_raw = 0 for v in videos[:max_v]: video_name = v.get("video_name") or v.get("video_nickname") or "" if not video_name: continue raw_path = os.path.join(args.raw_dir, video_name + ".json") if not os.path.isfile(raw_path): missing_raw += 1 continue raw = load_json(raw_path) raw_by_frame = raw_per_frame_objects(raw) all_count_diffs.extend(compare_counting(v, raw, video_name)) all_fr_diffs.extend(compare_frame_recall(v, raw_by_frame, video_name)) checked += 1 print(f"Part{args.part} 核查: 已对比 {checked} 个视频（原标注存在），{missing_raw} 个视频无原标注文件") print() print("=== object_counting 差异（benchmark answer vs 原标注同 concept 至末 clip 实例数）===") if not all_count_diffs: print(" 无差异") else: for video_name, concept, bench_answer, raw_count, _ in all_count_diffs[:50]: print(f" {video_name} | concept={concept} | benchmark={bench_answer} | raw={raw_count}") if len(all_count_diffs) > 50: print(f" ... 共 {len(all_count_diffs)} 条") print() print("=== frame_recall 差异（benchmark objects_in_frame vs 原标注该 frame_idx 的 objects[].concept）===") if not all_fr_diffs: print(" 无差异") else: for video_name, fidx, bench_objs, raw_objs, _ in all_fr_diffs[:30]: print(f" {video_name} frame_idx={fidx} | bench={bench_objs} | raw={raw_objs}") if len(all_fr_diffs) > 30: print(f" ... 共 {len(all_fr_diffs)} 条") print() print("Done.") if __name__ == "__main__": main()