Arko007/yoga-skeleton-rebuilt

Rebuilt Yoga Pose / Skeleton Dataset — preprocessing repository

Research-grade rebuild of a MediaPipe-derived yoga skeleton dataset from its original raw sources, with full provenance, quality control, and an independent audit. This repository reproduces the dataset from raw video; it does not patch the historical CSV.

Headline finding. The historical master_mlp_dataset_fully_classified.csv (654,488 rows) is not a filtered pose dataset. It is a per-frame dump of every decoded frame of the 12 source videos, in which ~83 % of rows carry MediaPipe output that the original pipeline's own report files admit was never actually detected. 99.36 % of its pose_label column is transition/unknown. Evidence and numbers are in docs/FINDINGS_historical_dataset.md.

---

1. What is in this repository

pipeline/yoga_core.py         canonical definitions: 33 landmarks, ST-GCN edges,
                              normalisation (verbatim from the notebook),
                              feature candidates, QC metrics
research/00_download_hf_videos.py   source-level dedup + verified download
research/00b_download_youtube.py   the 13 supplied YouTube sources
research/00b_audit_youtube.py       evidence of what could/could not be obtained
research/00c_source_registry.py     canonical source registry
research/01_historical_forensics.py what the old CSV actually contains
research/02_build_proxies.py        decode-once proxy generation
research/02b_smoke_mediapipe.py     MediaPipe sanity + throughput measurement
research/03_extract_landmarks.py    deterministic MediaPipe extraction + QC
research/03b_watcher.py             overlaps decode and inference
research/04_segment_and_label.py    segment manifest + rule-based pose labels
research/05_reverse_engineer_features.py  evidence-based feature definition search
research/06_sequences_and_split.py   ST-GCN tensors + source-level split
research/07_validate.py             the 12 independent audits
research/08_visual_verification.py  overlay panels + skeleton video previews
research/09_historical_comparison.py rebuilt vs historical agreement
research/10_stats_and_cards.py      dataset_statistics.json + DATASET_CARD.md

2. Source inventory (what actually exists)

	count	note
mp4 entries in the HF repo tree	24	the same 12 videos stored 3× each
unique raw videos (sha256)	12	6.83 GB, all SHA-256 verified
report JSONs in the repo	11	10 report 0 frames processed
supplied YouTube URLs	13	all identified via oEmbed
YouTube URLs already present on HF	1	Eml2xnoLpYE == HF 20201105_...
canonical sources after dedup	24	12 HF + 12 new YouTube
sources obtainable here	12	YouTube media is IP-blocked (see §7)

source_registry.csv carries source_id, source_type, source_url, source_filename, youtube_id, hf_paths_in_repo, hf_copy_count, duration_sec, fps, width, height, codec, sha256, sha256_verified, difficulty, status, duplicate_of, notes for all 25 rows.

3. Pipeline

00  inventory & dedup (by sha256, before any download)
00b acquire YouTube sources            00c source registry
01  forensics on the historical CSV
02  decode-once proxy  (10 fps, AR preserved, 640 long side)
03  MediaPipe Pose, 33 landmarks, model_complexity=1, deterministic
03b decode/inference overlap
04  segmentation + 15 features + rule-based labels
05  evidence-based reverse engineering of the 15 feature definitions
06  60-frame sequences + source-level split
07  twelve independent audits
08  visual verification (overlays + skeleton videos)
09  cross-validation against the historical CSV
10  statistics, dataset card, technical report

Run everything with:

bash run_all.sh          # idempotent; each stage skips completed work

4. Design decisions that matter

Proxy at 10 fps. The corpus is 4K/1080p AV1/VP9/H.264. 4K AV1 decodes at ~10 fps on this 2-core host, i.e. ~9 h just to read every frame once. We pay that cost exactly once into a proxy (config.yaml: proxy), then every later stage reads the cheap proxy. A 60-frame ST-GCN window therefore spans 6.0 s of practice. Aspect ratio is preserved — squashing 16:9 into 4:3 would silently distort every joint angle and therefore all 15 features.

Model complexity 1, not 0. Measured on 150 real yoga frames: the Lite model is only 22 % faster (13.5 vs 11.1 fps) but disagrees with the full model by a mean of 0.056 in normalised landmark units. Accuracy wins.

Normalisation is verbatim from the notebook. normalize_coordinate_sequence() in pipeline/yoga_core.py is a character-for-character reproduction of notebook cell 1 (pelvis = midpoint of landmarks 23/24, scaled by ‖hip_L − hip_R‖, 1e-5 guard). The ST-GCN edge list is likewise verbatim.

RAW and NORMALISED are both kept. stgcn_features_raw.npy is never overwritten by normalisation.

No frame-level train/test split. Sequences are grouped by source video; a source appears in exactly one split. Audit 11_split_leakage asserts it.

Non-overlapping windows. stride = 60 = window, so near-duplicate sequences are impossible by construction rather than filtered out afterwards.

5. Requirements

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

numpy<2 is required by the pinned mediapipe/jax stack. ENVIRONMENT.txt records the exact machine that produced the artefacts (2× Xeon 8163, 3 GB RAM, no GPU, ffmpeg 5.1.9).

6. Known deviations from the original pipeline

#	Deviation	Why	Documented in
1	10 fps proxy, 6 s windows	4K AV1 decode cost is ~9 h at native rate	config.yaml, dataset card
2	12 of 24 sources	YouTube media IP-blocked from this host	runs/run_001/youtube_acquisition_report.json
3	Frames with no detection are INVALID, not filled	the historical pipeline emitted hallucinated poses for them	docs/FINDINGS_historical_dataset.md
4	tree_pose rules corrected	the notebook's RULES['tree_pose'] defines knee_r/hip_r twice, so the first pair is silently overwritten by dict literal semantics	docs/FINDINGS_historical_dataset.md
5	15 feature definitions are reconstructed, not assumed	the historical extraction code was never published	runs/run_001/feature_reverse_engineering/

7. YouTube acquisition — honest status

youtube.com and its public oEmbed endpoint are reachable (all 13 titles resolved), but media extraction is blocked at the datacenter-IP level: yt-dlp returns "Sign in to confirm you're not a bot" for every player client (default, android, ios, tv_embedded, web_safari, mweb, android_vr), and no Piped/Invidious front-end returned a stream URL. The 12 genuinely new YouTube sources are therefore recorded as NOT_OBTAINED and excluded — they are not substituted with other videos. Eml2xnoLpYE is a duplicate of an HF file and was deliberately not downloaded twice. Full evidence: runs/run_001/youtube_acquisition_report.{json,csv}.

8. Traceability

Any sample in the released dataset can be traced back to its origin:

sequence_metadata.csv  -> source_id, subject_id, segment_id, start_frame,
                          end_frame, start_time, end_time, label, quality
segment_manifest.csv   -> source_id, segment_id, time range, segment_type, reason
source_registry.csv    -> source_id, source_url, youtube_id, sha256, provenance
raw landmarks .npz     -> per frame: x, y, z, visibility, verdict, QC metrics

9. Licence / ethics note

The raw videos are third-party YouTube content obtained from the dataset authors' Hugging Face mirror and are used here for academic research provenance verification. Redistribution of the raw video files is not authorised by this repository; the derived skeleton/label artefacts are the research output.