humanforall Stage-1: image → 595 2D LOD6 vertices
A Sapiens2 ViT fine-tuned to predict one heatmap per vertex of the MHR LOD6 body mesh (595 vertices),
for adults, children and infants. The decoded heatmap peaks are the 2D vertices, and the peak values are
a per-vertex confidence that Stage-2 uses (agelosk/humanforall-stage2). Code: agelosk/humanforall,
pose_stage1/ (inference: pose_stage1/infer_lod6.py).
Files
| path | backbone | trained on | use it? |
|---|---|---|---|
multipop_cross_v2_1b/ |
Sapiens2-1B | BEDLAM + AGORA + Anny-One (adult, child, infant) | yes: the current model (2026-09-28) |
multipop_cross/ |
Sapiens2-0.4B | the same blend | superseded by the 1B |
model.safetensors (repo root) |
Sapiens2-0.4B | BEDLAM only (adults) | only for the adult-only baseline |
Each folder holds model.safetensors and the Sapiens2 training config (config.py) it was trained with.
Results (multipop_cross_v2_1b)
Synthetic validation: every held-out person with a vertex in the image (673,936). PCKm@0.05 counts a vertex correct within 5% of the diagonal of the person's ground-truth 2D mesh box; the mean error (EPEn) is in the same units.
| group | n | PCKm@0.05 | mean error |
|---|---|---|---|
| adult | 620,312 | 0.932 | 2.0% |
| · BEDLAM | 116,674 | 0.955 | 1.7% |
| · AGORA | 7,811 | 0.904 | 2.3% |
| · Anny-One | 495,827 | 0.926 | 2.1% |
| child | 48,101 | 0.920 | 2.2% |
| · AGORA | 1,158 | 0.897 | 2.5% |
| · Anny-One | 46,943 | 0.920 | 2.2% |
| infant (Anny-One*) | 5,523 | 0.901 | 2.6% |
| all | 673,936 | 0.931 | 2.0% |
* Anny-One's "infants" are size-scaled small adults, not infant anatomy.
Real images, complete test sets, never trained on. PCKs@0.05 normalises by the extent of the valid ground-truth keypoints, which is stricter.
| source | n | PCKs@0.05 |
|---|---|---|
| MINI-RGBD (all 595 vertices; renders of real infant motion) | 10,000 frames | 0.979 |
| SyRIP (17 COCO keypoints, real infant photographs) | 500 | 0.671 |
| Relative Human (17 COCO keypoints) | 5,248 | 0.642 (adults 0.651, children 0.638, infants 0.592) |
The 17 keypoints are computed from the 595 vertices with a fixed linear map (r_coco17_lod6.npy in the
code repo). On real photographs a dedicated COCO keypoint model (ViTPose-H) scores higher (SyRIP 0.856,
Relative Human 0.795): the render-to-photograph gap is this model's main weakness.
Using it
- Crop with
infer_lod6.pyfrom code commit8abcd6for later: a square box around the person, padded 1.25×, not clipped to the image, warped with INTER_AREA when shrinking. Older code clipped the box and produced different (worse) 2D on people near the image edge. - The confidence is the raw heatmap peak per vertex (clamped at 0). Stage-2's current model needs it.
- All numbers above use ground-truth person boxes. Detector boxes are untested.
Gated: request access and it is granted manually.