3R-Adapter trains only 2.8M parameters: about 27% lower Chamfer distance than the strongest PEFT baseline in low-light 3D reconstruction
Synopsis
Feed-forward 3D reconstruction models degrade in long-tail settings such as low-light reconstruction, joint human-scene reconstruction and test-time adaptation, where cross-view attention becomes diffuse and correspondences fail. 3R-Adapter reframes adaptation as cross-view relation recovery: the backbone stays frozen and three lightweight modules handle deformable retrieval, relational rewiring and consistency refinement. On Oxford Day&Night with the CUT3R backbone, it uses 2.8M trainable parameters and lowers Chamfer distance by about 27% relative to the strongest PEFT baseline.
Interpretation
The failure of feed-forward 3D reconstruction models under domain shift is traced to collapse of cross-view relational structure, not to weak local feature representation. Earlier parameter-efficient fine-tuning methods such as LoRA and Adapter mainly perform channel-wise feature adjustment and rely on implicit attention updates for cross-view dependencies, giving limited control over relational routing. On Oxford Day&Night night images, the frozen model's Match@1 under exact token matching falls from 18.49% by day to 11.14% at night, confirming weaker cross-view matching.
Adaptation is recast as a relation-recovery problem carried out by three cooperating modules. Deformable Retrieval first predicts offsets and sampling weights for each query token within a sliding window over the view dimension, replacing dense all-to-all attention with sparse candidates. Relational Rewiring samples the candidates by bilinear interpolation and aggregates them with learned weights, and Consistency Refinement injects the result as a residual into frozen backbone features; on Oxford Day&Night with the CUT3R backbone, all three modules give a Chamfer distance of 0.581, retrieval alone 1.152, and retrieval plus rewiring 0.842.
On low-light reconstruction, 3R-Adapter reaches the best result among the compared PEFT methods with the fewest trainable parameters. On the CUT3R backbone it uses 2.8M trainable parameters and lowers Chamfer distance by about 27% relative to the strongest PEFT baseline. On the stronger permutation-equivariant backbone it uses 3.2M parameters and lowers Chamfer distance by about 32% relative to the original backbone; in test-time adaptation it lowers it by about 9% on 7-Scenes and about 7% on NRGBD relative to the strongest PEFT baseline.
Perspective
The demonstrated scope is low-light reconstruction on Oxford Day&Night, joint human-scene reconstruction on BEDLAM, and test-time adaptation on 7-Scenes and NRGBD, with DUSt3R, CUT3R and a permutation-equivariant model as backbones; adapters are trained on the corresponding training splits, and test-time adaptation optimizes each test scene separately with a self-supervised reconstruction loss.
The method still relies on the representations and geometric priors learned by the pre-trained backbone, reorganizing existing cross-view relations rather than learning reconstruction from scratch; when the backbone extracts no meaningful visual features or the input lies far outside the pre-training distribution, retrieval and rewiring may remain limited. The sliding window improves efficiency but may restrict long-range relation modeling in extremely long sequences or scenes with large viewpoint changes.
