qqyang/c2p-model

c2p — Coarse-to-Fine Contact → Pressure (Ego2Contact)

Bimanual tactile estimation from egocentric video, built on TouchAnything + EgoTouch. A decoupled two-stage design: Stage-1 trains a standalone contact head (per-cell contact probability p_c on the 21×21 grid, both hands) from a frozen DINOv2 view + WiLoR hand features; Stage-2/3 add pressure magnitude with a flow-matching GridDiT that consumes the Stage-1 contact map. This repo holds the checkpoints + full test-eval metrics for every stage.

Model families

dir	model	one-line
stage1_contact_head/	Stage-1 head	contact prob only, no magnitude (soft-BCE + Dice + part-CoP)
stage2a/	Stage-2a	frozen head (detached) → DiT-from-scratch, p_c as zero-init conditioning
stage2_gate/	Stage-2 GATE	frozen head p_c soft-gates the DiT output (floor 0.3), gate trained in-loop
stage3/	Stage-3 JOINT	init from GATE, unfreeze head + let the gated-final gradient co-train it (0.1× LR)

Each stageX/ has best_model.pth (+ last_model.pth for stage2a/stage3), and eval/stageX/test_{seen,unseen}_metrics.json. The two reference baselines' full evals (same protocol) are in eval/v2_dit/ and eval/ta_base/ (test_{seen,unseen}_metrics.json), for direct comparison with the c2p line.

Test eval — grid space, threshold @0.05, per-seq pooled (179 seen / 85 unseen)

Contact metrics for the pressure models are on the final pressure map; Stage-1 scores p_c>0.5 vs gt>0.05.

model	cIoU s	cIoU u	T.Acc s	T.Acc u	vIoU s	vIoU u	MAE↓ s	MAE↓ u	CoP↓ s	CoP↓ u
Stage-1 head	0.5117	0.4781	0.891	0.856	—	—	—	—	0.184	0.166
Stage-2a	0.5428	0.4647	0.889	0.835	0.442	0.315	0.042	0.061	—	—
Stage-2 GATE	0.5585	0.4739	0.898	0.838	0.479	0.325	0.039	0.060	0.164	0.155
Stage-3 JOINT	0.5680	0.4631	0.903	0.837	0.493	0.322	0.038	0.059	0.149	0.144
v2_dit (single-stage bar)	0.5735	0.4812	0.902	0.854	0.496	0.342	0.038	0.059	—	—
ta_base (regression bar)	0.5064	0.4720	0.890	0.838	0.425	0.345	0.049	0.064	—	—

Findings

• Stage-2a (conditioning) improves seen but its from-scratch DiT overfits seen magnitude and degrades unseen (0.4647, below the head) — the DiT re-learns localization from scratch.
• Stage-2 GATE (learned in-loop gate) is the best-balanced two-stage model: it fixes Stage-2a's OOD collapse (unseen +.009 and seen +.016 vs 2a) and beats the whole contact-prior family + a post-hoc late-fusion gate. It trails single-stage v2_dit but ties it on unseen.
• Stage-3 JOINT (unfreeze + co-train the head) gives the best in-distribution numbers (seen cIoU 0.568, edging v2_dit on seen T.Acc/MAE) and lifts the head's own contact cIoU 0.518→0.55, but sacrifices OOD — unseen drops to the family's lowest (0.4631). A clean demonstration that freezing the contact prior protects generalization.
• Cross-model failure analysis (per-task cIoU, GATE vs v2_dit vs ta_base): all three heads bottom out on the same tasks — press_multimeter_button, move_toolbox, fold_clothes_and_bag, shop_for_fruit, organize_kitchen_items, toss_and_catch_tennis_ball — in nearly the same rank order, so the long left tail is task-intrinsic (small / occluded / dynamic / OOD contact), not model-specific. The one sharp exception: on OOD move_toolbox the GATE is uniquely worst (0.326 vs v2_dit 0.361 / ta_base 0.360) — the frozen head under-localizes the occluded heavy-grip and the multiplicative gate enforces it, while the un-gated v2_dit and the area-over-predicting regression baseline both escape it. The concrete cost of the freeze.

Takeaway: the two-stage contact-prior closes most of the gap to a well-optimized single-stage flow model but does not overtake it; among two-stage variants, freeze-the-head (GATE) is the right default for OOD, while co-training (Stage-3) trades OOD robustness for in-distribution accuracy.

Demo videos (GT vs predicted pressure): qqyang/ego2contact-demo-videos.