Yoga-1M — AsanaAI skeleton dataset (RCC Institute of Information Technology, project P05)

1,206,391 frames from 24 yoga videos, stored as MediaPipe Pose landmarks (no video frames or images are redistributed here), with the 15 biomechanical joint angles used by the 3-head ResMLP and the 33-joint sequences used by the ST-GCN.

How the labels were made (and why they replace the older rule-derived labels)

A frame is a labelled pose only if ALL three hold: (1) the instructor named the pose (Whisper-large-v3 transcripts, English + Hindi, Sanskrit and Devanagari names), within [-1 s, +30 s]; (2) the app's own rule engine (rules_classifier.classify_pose with orientation cues) agrees the body is in that pose; (3) the pose is held steady (mean |angle change| < 8 deg over +-7 frames). Moving frames and steady frames with no recognised pose are transition/unknown; every ambiguous frame is __ignore__ (excluded, neither positive nor negative). Of the 1.21M frames: ~80k are cue-verified poses, ~421k transitions, ~705k excluded.

Contents (vol/ mirrors the working volume)

vol/csv/cue/cue_full.csv all frames (zero-z angles = what the deployed app feeds the model) · vol/csv/cue2/* MLP training sets (variants) · vol/cue per-video label arrays · vol/transcripts Whisper transcripts · vol/landmarks, vol/landmarks_new raw landmarks · vol/stgcn/cueT2_* ST-GCN windows (train + held-out-video test) · vol/photos real-photo landmark corpora (public HF sets + cleaned Commons/Openverse harvest) · vol/eval, vol/folds benchmark photos and folds · code/ the exact pipeline and the unmodified original trainers.

Evaluation protocol (what counts as a result)

Trainer-reported train/val accuracy is a random split and is NOT a result. Results are only: out-of-fold on 422 independent Commons photos, held-out public photos (incl. out-of-vocabulary poses), and held-out videos for the ST-GCN. Pass bar per pose: recall >= 0.70 AND precision >= 0.70, n >= 10. Models: https://huggingface.co/Arko007/asanaai-conference-runs