Research map / 3D vision / 3D generation
3R-Adapter: Retrieval, Rewiring, and Refinement for Efficient Adaptation of 3D Reconstruction Model
Zixiang Ai, Zhenyu Cui, Xu Zou, Jiahuan Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes 3R-Adapter, an adaptation framework (retrieval, rewiring, refinement) for feed-forward 3D reconstruction models to handle long-tail scenarios by addressing relational attention structure collapse.
Problem
Feed-forward 3D reconstruction models generalize well under large-scale pre-training but fail in long-tail scenarios such as low-light reconstruction, joint human-scene reconstruction, and test-time adaptation, attributed to collapse of cross-view attention correspondences.
Approach
Identifies relational attention structure collapse (failure of cross-view correspondences) as a primary cause and proposes an efficient adapter built around retrieval, rewiring, and refinement to restore that structure during adaptation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Relational Rewiring on top of Deformable Retrieval (ablation) | 3R-Adapter (CUT3R) · Oxford Day&Night | Chamfer: 1.152 -> 0.842 | Improved |
| add Deformable Retrieval alone (ablation) | CUT3R · Oxford Day&Night | Chamfer: 1.201 -> 1.152 | Improved |
| add Consistency Refinement alone (ablation) | CUT3R · Oxford Day&Night | Chamfer: 1.201 -> 1.012 | Improved |
| use all three components together (ablation) | CUT3R · Oxford Day&Night | Chamfer: 1.012 -> 0.581 | Improved |
| increase sampling locations K from 1 to 4 (ablation) | 3R-Adapter · Oxford Day&Night | Chamfer | Improved |
| use more than 4 sampling locations (ablation) | 3R-Adapter · Oxford Day&Night | Chamfer | No clear effect |
| increase relational fusion factor alpha beyond 0.10 (ablation) | 3R-Adapter · Oxford Day&Night | Chamfer | Hurt |
| overly large search radius rho (ablation) | 3R-Adapter · Oxford Day&Night | Chamfer | Hurt |
| replace channel-wise PEFT with 3R-Adapter | CUT3R · Oxford Day&Night | Chamfer: 0.791 -> 0.581 | Improved |
| adapt with 3R-Adapter | pi3 night Match@1 · Oxford Day&Night (block 31) | Match@1: 52.94 -> 52.16 | Mixed |
| adapt with 3R-Adapter | pi3 night Match@1 · Oxford Day&Night (block 21) | Match@1: 14.38 -> 19.21 | Improved |
| use 3R-Adapter vs specialized TTA methods | DUSt3R · 7-Scenes / NRGBD | Chamfer: 0.241 -> 0.216 (7-Scenes); 0.162 -> 0.205 (NRGBD) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 5 on most shared results.
- Human3R — Human3R: Everyone Everywhere All at Once (ICLR 2026) · better on 1 of 1 shared result
- Dark3R — Dark3R: Learning Structure from Motion in the Dark (CVPR 2026) · better on 1 of 1 shared result
- UniSH — UniSH: Unifying Scene and Human Reconstruction in a Feed-Forward Pass (CVPR 2026) · better on 0 of 1 shared result
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · better on 1 of 1 shared result
- DUSt3R — DUSt3R: Geometric 3D Vision Made Easy (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- H3R: Hybrid Multi-view Correspondence for Generalizable 3D Reconstruction (ICCV 2025 · similarity 0.89)
- View Transformation Robustness for Multi-View 3D Object Reconstruction with Reconstruction Error-Guided View Selection (AAAI 2025 · similarity 0.87)
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.87)
- The More You See in 2D the More You Perceive in 3D (CVPR 2024 · similarity 0.86)
- SIU3R: Simultaneous Scene Understanding and 3D Reconstruction Beyond Feature Alignment (NeurIPS 2025 · similarity 0.86)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
