CoolFace
Modelpublic

Cloth-splatters/dexgarmentlab-lift-20260822-state-est-gps-regression-tf2-noscale

sourceHugging Facemitupdated 11d agoView on Hugging Face
0likes
Model Card

DexGarmentLab Lift — GPS TF2 deterministic regression (D-GNN), no template-scale randomization

The deterministic ablation of the ClothAtlas paper ("D-GNN", Sec. V-A/V-B). The same GPSStateEstModel as the paper's flow estimator (dexgarmentlab-lift-20260822-state-est-gps-flow-tf2-noscale): sequential cross-attention, two point-cloud inputs, hidden 128, 8 layers / 8 heads, 2048-node cap, the same data and augmentation (joint scale 0.8-1.25 and yaw, template yaw up to 180 degrees with probability 0.5, no template-scale randomization, cloud corruption). Trained with the direct regression loss (Regression_StateEstGPS: the model sees [0 | rest] at timestep 0 and minimises a masked MSE on the vertex positions), so inference is one deterministic forward pass: --steps 1 --num-samples 1 is the complete evaluation. model/ holds the validation-selected EMA weights, config.yml the full training configuration.

  • —Training: job 1309183 (completed 2026-09-14, 9 h 34 min on 2 x GH200), 20 epochs / 398,700 optimiser updates, batch 2 x 16, learning rate 1e-3, 500 warm-up steps, cosine schedule, seed 259, bf16, EMA 0.9999, coord_scale 1.
  • —Dataset: dexgarmentlab_lift_full_state_20260822.h5, random-garment split, revision 771714a2; 176 / 23 / 22 train / validation / test garments.
  • —Best validation loss: 0.00013293644893884105. Weights SHA-256: f03de79a0af0503674d29c9034b88472589475bd5e0b2ba451f706819499a778.

Paper numbers (campaign paper_noscale_regression_20260914)

Exact-mesh configuration estimation over all 7,537 test frames (169 trajectories, stride 10), medians in mm, flow estimator in parentheses: surface Chamfer 12.8 (11.7), occluded surface Chamfer 13.0 (12.0), occluded worst decile 25.0 (22.9), vertex-error p90 36.1 (25.6).

Identification among 221 canonical meshes, 22 held-out garments, top-1 / top-5: flat frame 21 / 21 (22 / 22), lifted frame 8 / 14 (15 / 19), first three observations 20 / 21 (22 / 22), all nine 13 / 15 (18 / 22). Sequential identification with pruning (5 of 221 after 3 observations): exact mesh retained 21 / 22 (22 / 22), final top-1 12 / 22 (19 / 22).

Supersedes dexgarmentlab-lift-20260822-state-est-gps-regression-tf2 (10 epochs, template-scale randomization) for the submitted paper.

Loading

python
from src.hub import resolve_checkpoint
from src.inference.checkpoint import load_pipeline

pipe = load_pipeline("dexgarmentlab-lift-20260822-state-est-gps-regression-tf2-noscale")

Code: https://github.com/jsll/UniClothDiff/tree/icra