notesum.ai
Published at December 6DenseMatcher: Learning 3D Semantic Correspondence for Category-Level Manipulation from a Single Demo
cs.RO
cs.CV
Released Date: December 6, 2024
Authors: Junzhe Zhu1, Yuanchen Ju2, Junyi Zhang3, Muhan Wang4, Zhecheng Yuan5, Kaizhe Hu5, Huazhe Xu5
Aff.: 1Tepan Inc.; 2Shanghai Qi Zhi Institute; 3UC Berkeley; 4Stanford University; 5Shanghai AI Lab

| All | Held-out | |||
| Methods | AUC | Err | AUC | Err |
| ConsistFMap (FAUST) (Cao & Bernard, 2022) | 0.537 | 7.86 | 0.497 | 8.39 |
| ConsistFMap (DenseCorr3D) (Cao & Bernard, 2022) | 0.541 | 7.23 | 0.502 | 7.92 |
| URSSM (FAUST) (Cao et al., 2023) | 0.568 | 6.37 | 0.532 | 7.07 |
| URSSM (DenseCorr3D) (Cao et al., 2023) | 0.589 | 6.08 | 0.539 | 6.87 |
| Diff3F (Dutt et al., 2024) | 0.522 | 5.96 | 0.423 | 8.53 |
| DenseMatcher (Ours) | 0.845 | 1.74 | 0.775 | 2.82 |
| w/o DiffusionNet | 0.672 | 4.74 | 0.662 | 5.53 |
| w/o Preservation Loss | 0.568 | 5.11 | 0.509 | 6.92 |
| w/o FeatUp | 0.741 | 3.48 | 0.638 | 5.78 |
| w/o Constraint for FMap | 0.824 | 1.98 | 0.735 | 3.32 |