Verified methods
4
Every integrated method has metrics and video
Monocular Motion Capture · 3DPW Test
Camera-relative human motion recovery from monocular video. Every integrated method has a public video demo, while metric rows retain their native input protocol and body-model disclosure.
3DPW official test
Camera-space evaluation
SMPL / SMPL-X / SOMA
Verified coverage
Native method videos
Current public snapshotLoading verified results
Target-crop methods
3
Official target tracks
Full-frame methods
1
Native multi-person inference
Public video demos
4
Native body-model output
Method comparison
Diagnostic comparison with input protocol shown explicitly; no cross-protocol formal rank is assigned.
Official target cropFull-frame nativeGT reference
PA-MPJPE
Lower is better; chart colors identify input protocol.
Normalized diagnostic profile
100 is the best displayed value per axis; this is not a shared formal rank.
All-method visual comparison
One browser-playable native output for every integrated monocular-capture method.
Verified results · 3DPW camera
GT is pinned first. Target-crop and full-frame rows remain visibly separated by their input disclosure.
| Method | Input protocol | Coverage | PA-MPJPE ↓ | MPJPE ↓ | PVE ↓ | Accel. error ↓ |
|---|
Joint and vertex errors are in mm; acceleration error is in m/s².PVE is omitted where predicted and GT body-model surfaces are incompatible.
Protocol details
3DPW camera metrics
Official person tracks and valid masks. Pelvis-aligned MPJPE, per-frame similarity-aligned PA-MPJPE, compatible-surface PVE, and fps-aware acceleration.
Input disclosure
GVHMR, GEM-SMPL, and GEM-X use official target crops. PromptHMR-Video uses its native full-frame multi-person pipeline.
Cross-model contract
SMPL, SMPL-X, and SOMA comparisons use explicitly named common joints. Missing compatible surfaces remain unavailable rather than fabricated.