momo321654/RepCountA-videos

RepCount-A — playable MP4 package

RepCount Part-A (LLSP) repackaged as plain H.264 MP4 files plus per-split JSON annotations, for direct use in video-LLM evaluation pipelines and standard players.

• 1,041 videos — train 758 / validation 131 / test 152 (~9 GB, mostly 360p, native aspect)
• Fine-grained annotations: repetition count + per-cycle boundary frame indices (cycle_bounds)

Structure

train/                       758 .mp4
validation/                  131 .mp4
test/                        152 .mp4
train_annotations.json
validation_annotations.json
test_annotations.json

Each annotation entry:

{
  "video": "stu10_0.mp4",
  "video_id": "stu10_0",
  "source_name": "stu10_0.mp4",
  "action_type": "pull_up",
  "count": 14,
  "cycle_bounds": [1, 23, 23, 40, ...]
}

cycle_bounds is a flat list of frame indices marking the start/end of each action cycle (pairs), as annotated by the original authors. Standard metrics: MAE and OBO (off-by-one accuracy).

Provenance

Videos and annotations originate from the official RepCount release (RepCountA.tar.gz, LLSP structure), obtained via the lmms-lab-eval/repcounta-lance mirror; video bytes are extracted as-is (no re-encoding). Annotation frame indices were spot-verified against decoded frame counts.

License / attribution

RepCount is released for academic use by its authors. If you use this data, cite:

@inproceedings{hu2022transrac,
  title={TransRAC: Encoding Multi-scale Temporal Correlation with Transformers for Repetitive Action Counting},
  author={Hu, Huazhang and Dong, Sixun and Zhao, Yiqun and Lian, Dongze and Li, Zhengxin and Gao, Shenghua},
  booktitle={CVPR},
  year={2022}
}

Official dataset page: https://svip-lab.github.io/dataset/RepCount_dataset.html