3R-Adapter: Retrieval, Rewiring, and Refinement for Efficient Adaptation of 3D Reconstruction Model
Organizations: Peking University · Tsinghua University · Huazhong University of Science and Technology
Abstract
Feed-forward 3D reconstruction models have demonstrated strong generalization under large-scale pre-training, yet they struggle in long-tail scenarios such as low-light reconstruction, joint human-scene reconstruction, and test-time adaptation. We observe relational attention structure collapse, manifested as the failure of cross-view correspondences, as a primary cause of these limitations. While adaptation offers a practical way to bridge this gap, existing parameter-efficient fine-tuning (PEFT) approaches mainly perform channel-wise feature adjustment and lack mechanisms to explicitly adapt relational structures, leaving correspondence failures unresolved. In this work, we formulate adaptation of 3D reconstruction models as a problem of correspondence recovery under domain shift and propose 3R-Adapter, a PEFT paradigm for long-tail 3D reconstruction scenarios. Our method consists of three components: Deformable Retrieval for task-adaptive relation candidate retrieval, Relational Rewiring for reconstructing cross-view relational structures, and Consistency Refinement for injecting the rewired relations into stable predictions. Extensive experiments on low-light reconstruction, joint human-scene reconstruction, and test-time adaptation demonstrate that our approach significantly improves robustness over existing PEFT methods while maintaining efficient lightweight adaptation. Our code is released at https://github.com/PKU-OV3-LAB/3R-Adapter.
Figures & tables
| Method | Publication | Params. | Camera Pose Estimation | Scene Reconstruction | |||||
| ATE | RPE-T | RPE-R | Accuracy | Complete | Chamfer | Normal | |||
| DUSt3R [ 1 ] | CVPR 24 | 0.57B | 0.937 | 0.729 | 10.351 | 0.741 | 0.598 | 1.339 | 0.526 |
| CUT3R [ 2 ] | CVPR 25 | 0.79B | 0.784 | 0.542 | 9.924 | 0.562 | 0.639 | 1.201 | 0.534 |
| VGGT [ 3 ] | CVPR 25 | 1.19B | 0.401 | 0.324 | 4.236 | 0.356 | 0.585 | 0.941 | 0.522 |
| Dark3R [ 13 ] | CVPR 26 | 0.69B | 0.387 | 0.285 | 4.019 | 0.473 | 0.424 | 0.897 | 0.532 |
| [ 4 ] | ICLR 26 | 0.96B | 0.345 | 0.261 | 3.657 | 0.190 | 0.746 | 0.936 | 0.518 |
| Method | Publication | Params. | 7-Scenes [ 40 ] | NRGBD [ 41 ] | ||||||
| Accuracy | Complete | Chamfer | Normal | Accuracy | Complete | Chamfer | Normal | |||
| DUSt3R [ 1 ] | CVPR 24 | 0.57B | 0.146 | 0.181 | 0.327 | 0.736 | 0.144 | 0.154 | 0.299 | 0.871 |
| MASt3R [ 29 ] | CVPR 24 | 0.63B | 0.189 | 0.211 | 0.400 | 0.687 | 0.085 | 0.063 | 0.148 | 0.794 |
| CUT3R [ 2 ] | CVPR 25 | 0.79B | 0.126 | 0.154 | 0.280 | 0.727 | 0.099 | 0.076 | 0.175 | 0.837 |
| Test3R [ 17 ] | NeurIPS 25 | 0.8M | 0.105 | 0.136 | 0.241 | 0.746 | 0.083 | 0.079 | 0.162 | 0.870 |
| Online3R [ 18 ] | CVPR 26 | 1.2M | 0.112 | 0.145 | 0.257 | 0.733 | 0.084 | 0.081 | 0.165 | 0.829 |
| Method | Publication | Params. | Scene Reconstruction | Human Mesh Reconstruction | |||||
| Accuracy | Complete | Chamfer | PA-PVE | Metric-PVE | PA-MPJPE | Metric-MPJPE | |||
| DUST3R [ 1 ] | CVPR 24 | 0.57B | 0.237 | 0.159 | 0.396 | - | - | - | - |
| CUT3R [ 2 ] | CVPR 25 | 0.79B | 0.185 | 0.096 | 0.281 | - | - | - | - |
| VGGT [ 3 ] | CVPR 25 | 1.19B | 0.092 | 0.044 | 0.136 | - | - | - | - |
| Joint Human-Scene Reconstruction | |||||||||
| JOSH3R [ 16 ] | ICLR 26 | 0.66B | 0.231 | 0.107 | 0.338 | 46.2 | 245.7 | 41.5 | 239.4 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Deformable Retrieval | Relational Rewiring | Consistency Refinement | Accuracy | Complete | Chamfer | Normal |
| - | - | - | 0.562 | 0.639 | 1.201 | 0.534 |
| ✓ | - | - | 0.540 | 0.612 | 1.152 | 0.534 |
| ✓ | ✓ | - | 0.397 | 0.445 | 0.842 | 0.542 |
| - | - | ✓ | 0.483 | 0.529 | 1.012 | 0.536 |
| ✓ | ✓ | ✓ | 0.254 | 0.327 | 0.581 | 0.550 |
3D Reconstruction
3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.
See all 1,025 papers