Datasets:
metadata
license: other
language:
- en
- zh
pretty_name: dance-data (curated OmniDance annotations + videos)
tags:
- dance
- motion-caption
- motionscript
dance-data
Private-collection dance annotation data derived from the curated OmniDance set (222,939 source clips; music + reference dance video per clip).
Contents
| path | description |
|---|---|
| manifests/ | Curated 222,939-source manifest summary + per-batch manifests; original-MS export manifest (221,887 sources, SHA256 per pack/member) |
| ms_original_planner_dataset_v1/ | Planner training JSONL (train/val/test, per-source records with dance style + timed MotionScript target) |
| raw_motionscript_packs/ | 5,849 tar packs, 50 sources each: original MotionScript draft (Chinese 2D-pose rule sentences), full timed events, merged-edge bar window plan, rich-only dance-style candidates. NOT VLM-rewritten. |
| raw_rich_captions/ | Original rich five-chapter captions (317k .txt packed into 48 tars) |
| style_labels/ | Historical VLM dance-style labels (22-genre vocabulary + "unknown") for 11,992 sources |
| stage2_vlm_revised/ | VLM visual audit + fused caption for ~7,840 sources |
| videos_tars/ | All 263,350 source videos (704x1280 mp4, ~5s clips) packed 5,000 per tar, uncompressed |
| audio_tars/ | All 263,350 paired audio wavs packed 5,000 per tar |
Format notes
- Target caption format (planner): first line "Dance style: <one of 22 genres or unknown>", then per merged bar window "[start-end s] description" (e.g. "[0.000-1.960s] ..."). Windows come from BeatThis downbeats with short-edge merging; raw bar boundaries preserved. Window times indicate provenance only, not action boundaries.
- Genre vocabulary: chinese classical, chinese folk, ballet, contemporary, lyrical, modern jazz, jazz funk / heels, k-pop, hip-hop, popping, locking, breaking, house, krump, waacking, street jazz, latin, belly dance, indian / bollywood, other ethnic / folk, pop / casual, freestyle / other; "unknown" when undetermined.
- Known limitations: 2D-pose side wording does not guarantee anatomical sides; last ~0.06s of each clip unobserved; beat/downbeat detections and audio-video alignment unverified; style labels are model-derived, not human ground truth. All annotation outputs are unreviewed and not training-approved.
License
To be determined.