Benchmark 直接检查结果
检查时间:已运行脚本完成。
数据源:本地 clone 自 HF
- Part1:
data/hf_ready4label/new_long_video/corrected_json_2(对应 HFTorwnexial/ready4label→new_long_video/corrected_json_2) - Part2+3:
data/hf_ready4label/top20merge_full/corrected_json_2(对应 HFtop20merge_full/corrected_json_2)
1. Part1(长视频,40 个视频)
| 检查项 | 结果 | 说明 |
|---|---|---|
| 视频覆盖 | 40/40 | 所有 benchmark 视频在原标注中均有对应 JSON |
| frame_recall | 通过 | benchmark 中每帧的 objects_in_frame 与原标注该 frame_idx 的 objects[].concept 一致,无差异 |
| object_counting | 口径不同 | benchmark 的 answer(如 131、303)与「原标注到末 clip 的检测框数累加」不是同一口径(raw 为数千级,benchmark 为百级)。benchmark 应为人工或唯一实例数,不是框数累加,故 110 条均显示差异属预期 |
结论:帧级物体实例(frame_recall)与原标注完全对应;counting 为另一套定义,无法用当前 raw 自动核对。
2. Part2(短视频 Place&Motion,424 视频)
| 检查项 | 结果 | 说明 |
|---|---|---|
| 视频覆盖 | 401 有原标注,25 无 | 25 个视频在原标注目录中无 JSON(执行时会被跳过) |
| object_counting | 通过 | 有 counting 的 benchmark 与「原标注到末 clip 的框数」一致,无差异 |
| frame_recall | 通过 | benchmark 的 objects_in_frame 与原标注该 frame 的 objects[].concept 一致,无差异 |
结论:Part2 标注与原标注一致(在可比口径下)。
3. Part3(短视频 Objects,382 个视频)
| 检查项 | 结果 | 说明 |
|---|---|---|
| 视频覆盖 | 382/382 | 所有 benchmark 视频在原标注中均有对应 JSON |
| frame_recall | 通过 | benchmark 的 objects_in_frame 与原标注一致,无差异 |
| object_counting | 口径不同 | 1329 条显示差异:benchmark 数值与「原标注框数累加」口径不同(有人工/唯一实例等),与 Part1 同理 |
结论:帧级物体实例(frame_recall)与原标注完全对应;counting 为另一套定义。
4. 汇总
- frame_recall(帧内物体列表):Part1、Part2、Part3 均与原标注一致,无差异。
- object_counting:Part2 在「到末 clip 框数」口径下一致;Part1、Part3 的 benchmark 与 raw 框累加口径不同,差异来自定义不同,非标注错位。
- 视频列表:Part1 全对齐;Part2 有 25 个仅 benchmark、11 个仅原标注(你说执行时会跳过);Part3 为原标注子集,正常。
总体:标注文件内的物体实例(帧级)与原标注对应;counting 需以 benchmark 当前定义(人工/唯一实例等)为准,无法用 raw 框数自动校核。