dexgarmentlab-lift-20260822-retrieval-embedding

DexGarmentLab Lift — learned retrieval embedding (LRE baseline)

The ClothAtlas paper's learned identification baseline (LRE, Sec. V-C). RetrievalEmbedding (src/models/retrieval_embedding.py): one shared PointNet++ trunk embeds a partial cloud and the surface samples of a canonical rest mesh into 256 dimensions; candidates are ranked by 1 - cosine similarity.

• Training: scripts/train_retrieval_embedding.py, job 1153551 (completed 2026-09-08, 12 min on one GH200). Each batch pairs one random frame (stride 5, the estimator's cloud corruption, joint scale 0.8-1.25 and yaw) of 64 distinct training garments with their rest meshes; symmetric InfoNCE at a fixed temperature 0.07, dropout 0.3, AdamW 1e-4 / weight decay 0.1, 6 epochs of 500 steps, one warm-up epoch, cosine schedule, seed 0, validation every 250 steps against the 221-mesh library.
• Best checkpoint at epoch 2, step 1250: validation first-frame top-1 17.2 %, top-5 37.0 % (23 validation garments). Weights SHA-256: ab0395dab9c74fbae32a85e79b3de89686fa3150914a5a8699bad46dc1ad2938.

Paper numbers

Identification among 221 canonical meshes, 22 held-out garments, top-1 / top-5: flat frame 0 / 5, lifted frame 1 / 4, first three observations 0 / 6, all nine 1 / 5 (ClothAtlas: 22 / 22, 15 / 19, 22 / 22, 18 / 22).

Loading

from src.hub import resolve_checkpoint
from src.models.retrieval_embedding import RetrievalEmbedding

model = RetrievalEmbedding.load(resolve_checkpoint("dexgarmentlab-lift-20260822-retrieval-embedding"))

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