dexgarmentlab-lift-20260822-point-completion

DexGarmentLab Lift — point-cloud completion network (CR baseline)

The completion half of the ClothAtlas paper's "completion + registration" baseline (CR, Sec. V-B). PointCompletionNet (src/models/point_completion.py): a PointNet++ encoder and a PCN-style decoder that emits 512 coarse points and a 2 x 2 folding grid around each of them (2048 fine points), trained with the symmetric L1 Chamfer distance on both outputs.

• Input: one partial cloud with the estimator's training corruption (jitter, random and wedge dropout, outliers), jointly scaled (0.8-1.25) and yawed with its target, centred and resampled to 2048 points with a validity mask. Target: 2048 samples on the ground-truth surface in the same frame.
• Training: scripts/train_point_completion.py, job 1152819 (completed 2026-09-08, 1 h 44 min on one GH200): 128,136 training frames (stride 5), 20 epochs, batch 64, AdamW 3e-4 / weight decay 0.05, one warm-up epoch, cosine schedule, bf16 autocast, seed 0.
• Best epoch 18: validation fine Chamfer median 13.33 mm (coarse 20.13 mm) on 3,535 validation frames. Weights SHA-256: d67ecfcc63cbb68dd596c5a852cfe690ab498f0f02e8676436a1ae31306a7363.

Paper numbers

CR = this network followed by non-rigid registration of the canonical mesh to the completed cloud (scripts/eval_reconstruction_completion.py, edge-strain weight 0, Laplacian weight 100, chosen on the validation split). Over all 7,537 test frames: surface Chamfer 16.0 mm, occluded 16.0 mm, occluded worst decile 37.4 mm, vertex-error p90 332 mm (the registration recovers the surface, not the correspondence).

Loading

from src.hub import resolve_checkpoint
from src.models.point_completion import PointCompletionNet

net = PointCompletionNet.load(resolve_checkpoint("dexgarmentlab-lift-20260822-point-completion"))

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