Human Motion & Pose
Culturally diverse human movement — dance, martial arts, sport — with whole-body 2D pose, hands, face and a 3D body reconstruction on every frame.
See the annotation, live
Real clips with annotation JSON rendered live over the video. Toggle the overlay off to compare against the raw clip — the difference is the product.
Interactive samples for this line are being prepared. The record schema on this page shows exactly what each record contains.
Two layers, scoped separately
The curated base layer is always included. Dense annotation is an optional add-on on the same clips — take depth only where your training recipe needs it.
- Style / genre, action label and culture-region tag per clip
- Person count, camera motion, motion-strength, full-body and truncation fractions, occlusion flags
- Shot-integrity verdict per clip: no hidden cuts inside a delivered clip
- Rights status and a ready-to-paste attribution line per record
- 2D whole-body pose: 133 COCO-WholeBody keypoints per person per frame — body, feet, 68 face points, two 21-point hands — with track ids and per-frame validity
- 3D body: MHR (Momentum Human Rig) joints, shape and per-frame camera — projection agrees with the 2D layer to sub-pixel
- Per-frame hand-reliability flag from an independent second pass, measured against blind golden sets — shipped as data, and drawn live in the viewer
- Per-clip overlay proof video: 2D, second track and projected 3D rendered over the frames
Machine-readable contract: motion.schema.json
Measured on golden sets
Numbers below come from real annotation runs measured against human golden labels — with sources, not adjectives.
| Metric | Value | Context | Source |
|---|---|---|---|
| Licence verification | 106 / 110 | candidates whose licence string matched byte-for-byte at download time; 30/30 masters re-verified with zero drift | sourcing report, August 2026 |
| Package acceptance | 74 + 24 | build-time checks plus an independent verifier over the shipped files — zero failures | public demo delivery, August 2026 |
| Hand-flag residual | 3.62% | measured share of fabricated points among flag-confirmed hands (95% CI 1.39–6.33%); the flag is per frame and per hand — the viewer above draws unconfirmed hands dashed, so you can watch exactly where it fires | two rounds of blind golden-set probes |
| 2D ↔ 3D consistency | 1e-4 px | reprojection of the 3D body against the model's own 2D keypoints, verified at acceptance | delivery acceptance run |
What we tell you before you ask
- Hand keypoints are shown, not silently trusted. Their measured reliability flag misses our ≤5% bar at the upper end of its confidence interval, and on this demo set it fires on about a third of all hands (15–67% depending on the clip) — the viewer draws every one of those dashed instead of hiding the problem. The evidence pack to recompute our numbers without our code ships in the delivery.
- The hand flag is the only per-keypoint quality signal in this delivery: face points are shipped unmeasured, and the flag itself is measured per source frame but published on the viewer's ~12 Hz sample grid, so a burst shorter than a sample can fall between two published verdicts.
- The 3D body is a per-frame model reconstruction (MHR) accepted by visual QA — it is not motion capture.
- The public demo set is 8 clips across 8 styles; production volumes, style balance and annotation depth are scoped per brief.
- Acceptance removed data rather than waving it through: two culture buckets planned for this set were dropped when the reviewers found chimeric tracks, and 10 planned clips became 8. A third of the GPU budget for the line went into clips that were then rejected.
Sourced from the open public web, curated for motion clarity and cultural diversity — single-shot clips, verified full-body presence, license string checked byte-for-byte at download time. The public demo set is 100% CC BY 3.0 with per-clip attribution shipped in the package; production batches document rights per record.
License class: CC BY 3.0 (demo set), documented per record. Per-record provenance — source, license, fetch time, transform hash — ships with every delivery. Methodology →
- WebDataset shard + per-clip JSON records
- CSV / Parquet index
- Overlay proof video per clip
- numpy-only loader (Apache-2.0)
Loader snippets for HuggingFace datasets, WebDataset and
PyTorch are in the docs.
Three steps
- 01 Explore the public samples
Right on this page — schema, QA numbers and live viewer where cleared.
- 02 Request an eval batch
A private sample matched to your domain, under NDA, usually within 48 hours.
- 03 License the batch or scope the build
Take the current batch as it stands, or scope volume, annotation depth, license constraints and refresh cadence to your training recipe.