shriyan2407777/infiniterep-physiotherapy

InfiniteRep — Physiotherapy Exercise Archive (rescued)

Provenance. InfiniteRep is a large synthetic human-exercise dataset originally released by Infinity AI. As Infinity AI is no longer operating and the original distribution is unavailable, this repository preserves the archive and publishes the exact instance-disjoint, class-stratified train / dev / test manifests used in the PG-XFormer study. Full credit for the underlying synthetic data belongs to the original Infinity AI authors; this is a research-use archival mirror with attribution.

Contents

• —train.parquet, dev.parquet, test.parquet — per-instance metadata (exercise class, instance id, frame count, keypoint coverage, archive path, size).
• —metadata_full.parquet — the complete scan before splitting.
• —eda/ — exploratory tables and figures.
• —data/<split>/<class>/*_img_labels.zip — raw per-instance frame + keypoint archives.

Splits

Disjoint by instance and stratified so every class appears in every split. dev and test are held out and were fixed before any model selection.

Classes (11): armraise (70), bicyclecrunch (70), birddog (70), curl (70), data (200), fly (70), legraise (70), overheadpress (70), pushup (70), squat (70), superman (70)

Intended use & role

Synthetic pretraining corpus for sim-to-real physiotherapy exercise recognition. In the PG-XFormer protocol the model is pretrained here, then fine-tuned and evaluated on clinical KIMORE and tested externally on UI-PRMD. This is not a clinical dataset and contains no patient data.

Citation

Please credit the original Infinity AI InfiniteRep dataset and cite the PG-XFormer paper (see the GitHub repository).

License

Research use with attribution to Infinity AI. If you are a rights holder and wish this archival mirror amended, please open an issue on the GitHub repository.