JointHand: per-vertex hand contact and hand–object distance

jointhand_head.pt is the JointHand contact head (42.4 M parameters, epoch 29). It runs on top of the frozen WiLoR backbone (wilor_final.ckpt, unchanged), so the hand mesh is exactly WiLoR's. For every one of the 778 MANO vertices it predicts

• contact_prob: contact probability (paper read-out: p > 0.52)
• pred_dist: metric hand–object distance in metres (paper read-out: d < 12 mm)

The default read-out on datasets with distance supervision is contact OR distance.

file	contents
jointhand_head.pt	{'contact_head_state_dict', 'epoch', 'head', 'head_dims', 'env'}

The head reads these environment variables at construction time; they must match training (they are also stored under env in the checkpoint):

HACO_DCC_V2_XYZ=1  HACO_DCC_V2_DEEP_DIST=1  HACO_DIST_LOG=1  HACO_DIST_INJECT_TAU=0.015  HACO_FORCE_INJECT_P0=0.10

Architecture: AttentionSamplingDistForceInjectContactHead(vit_feat_dim=1280, model_dim=512, num_heads=8, num_layers=4, mlp_dim=2048, dropout=0.1). The paper's numbers were computed under bf16 autocast.

• Demo: https://huggingface.co/spaces/Xgxx00200/JointHand
• Project page / paper / code: see the demo page

The pressure branch of this checkpoint is not trained to a useful output and should not be used.

Licence: CC BY-NC 4.0 (research use). WiLoR and MANO keep their own licences.