Research map / 3D vision / Depth and stereo
UMIFormer: Mining the Correlations between Similar Tokens for Multi-View 3D Reconstruction
Zhenwei Zhu, Liying Yang, Ning Li, Chaohao Jiang, Yanyan Liang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add IVDB (inter-view token rectification) with STM (ablation) | UMIFormer · ShapeNet (3 views) | IoU: 0.7477 -> 0.7518 | Improved |
| use STM instead of pooling-based merger (with IVDB) (ablation) | UMIFormer · ShapeNet (3 views) | IoU: 0.7372 -> 0.7518 | Improved |
| use STM instead of attention-based merger (with IVDB) (ablation) | UMIFormer · ShapeNet (20 views) | IoU: 0.7593 -> 0.7702 | Improved |
| rectify tokens by predicting offset rather than FC mapping directly (ablation) | IVDB · ShapeNet (3 views) | IoU: 0.6935 -> 0.7518 | Improved |
| predict offset vs offset weight in IVDB (ablation) | IVDB · ShapeNet (3 vs 20 views) | IoU: 3 views 0.7546 vs 0.7518; 20 views 0.7694 vs 0.7702 | Mixed |
| joint attention over all views' tokens instead of independent branches (ablation) | ViT encoder · ShapeNet (20 views) | IoU: 0.7566 -> 0.7252 | Hurt |
| video-transformer factorised decoupling instead of similar-token IVDB (ablation) | ViT encoder · ShapeNet (3 views) | IoU: 0.7394 -> 0.7400 (Factorised Attention) | No clear effect |
| insert IVDB repeatedly rather than once (ablation) | UMIFormer · ShapeNet (8 views) | IoU: 0.7606 -> 0.7661 | Improved |
| initialize intra-view blocks from pre-trained ViT (ablation) | UMIFormer · ShapeNet (20 views) | IoU: 0.5082 -> 0.7702 | Improved |
| train UMIFormer+ with 8 views instead of 3 views | UMIFormer · ShapeNet (1 view) | IoU: 1 view 0.6802 -> 0.5672; 20 views 0.7702 -> 0.7886 | Mixed |
| use IVDB and STM with other decoders (EVolT, LegoFormer) (ablation) | UMIFormer · ShapeNet (3 views, EVolT decoder) | IoU: 0.7399 -> 0.7536 | Improved |
Later papers that compared with this method
1 later accepted paper reports UMIFormer as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Long-Range Grouping Transformer for Multi-View 3D Reconstruction (ICCV 2023 · better on 5 of 7 shared results)
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
- Uni3R: Unified 3D Reconstruction and Semantic Understanding via Generalizable Gaussian Splatting from Unposed Multi-View Images (CVPR 2026 · similarity 0.87)
- Share with Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency (ECCV 2022 · similarity 0.85)
- Multi-view Pyramid Transformer: Look Coarser to See Broader (CVPR 2026 · similarity 0.85)
- NU-MCC: Multiview Compressive Coding with Neighborhood Decoder and Repulsive UDF (NeurIPS 2023 · similarity 0.85)
- Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers (NeurIPS 2026 · similarity 0.84)
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.84)
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.
