Research map / Architectures and efficiency / Attention and transformers
CoMatch: Dynamic Covisibility-Aware Transformer for Bilateral Subpixel-Level Semi-Dense Image Matching
Zizhuo Li, Yifan Lu, Linfeng Tang, Shihua Zhang, Jiayi Ma
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 CGTC token condenser (vs depth-wise conv + max pooling) (ablation) | CoMatch · ScanNet | AUC@5deg: 20.0 -> 21.8 | Improved |
| use CAA instead of vanilla attention (ablation) | CoMatch · ScanNet | AUC@5deg: 20.7 -> 21.8 | Improved |
| use CAA instead of flash attention (ablation) | CoMatch · ScanNet | AUC@5deg: 20.5 -> 21.8 | Improved |
| use CAA instead of linear attention (ablation) | CoMatch · ScanNet | AUC@5deg: 18.1 -> 21.8 | Improved |
| use proposed 2nd-stage bilateral refinement instead of LoFTR's refinement (ablation) | CoMatch · ScanNet | AUC@5deg: 20.4 -> 21.8 | Improved |
| add 2nd-stage refinement (ablation) | CoMatch · ScanNet | AUC@5deg: 20.2 -> 21.8 | Improved |
| add the BSR module (ablation) | CoMatch · ScanNet | AUC@5deg: 11.3 -> 21.8 | Improved |
| use proposed backbone instead of RepVGG (ablation) | CoMatch · ScanNet | AUC@5deg: 21.2 -> 21.8 | No clear effect |
| use proposed backbone instead of ResNet-FPN (ablation) | CoMatch · ScanNet | AUC@5deg: 21.5 -> 21.8 | No clear effect |
| condensing range 2x2 instead of 4x4 (ablation) | CoMatch · MegaDepth-1500 | runtime (ms): 123.8 -> 207.6 | Hurt |
| increase test resolution from 800x800 to 960x960 | CoMatch · MegaDepth | AUC@5deg: 55.5 -> 56.3 | Improved |
| use CoMatch instead of ASpanFormer | semi-dense matcher · MegaDepth-1500 | runtime (ms): 211.8 -> 123.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ASpanFormer — ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer (ECCV 2022) · better on 4 of 4 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
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.87)
- ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer (ECCV 2022 · similarity 0.87)
- Raising the Ceiling: Conflict-Free Local Feature Matching with Dynamic View Switching (ECCV 2024 · similarity 0.86)
- CoMapGS: Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis (CVPR 2025 · similarity 0.84)
- Dynamic Cues-Assisted Transformer for Robust Point Cloud Registration (CVPR 2024 · similarity 0.84)
- Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence (ICLR 2024 · similarity 0.83)
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.
