Research map / Architectures and efficiency / Attention and transformers
ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer
Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, Long Quan
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 hierarchical global-local attention with fixed 13 px span over global-only attention (ablation) | ASpanFormer · ScanNet | pose AUC@5deg: 22.65 -> 24.85 | Improved |
| make local attention span adaptive from flow uncertainty instead of fixed 13 px (ablation) | ASpanFormer · ScanNet | pose AUC@5deg: 24.85 -> 25.61 | Improved |
| make local attention span adaptive (AUC@20) (ablation) | ASpanFormer · ScanNet | pose AUC@20deg: 62.71 -> 63.33 | Improved |
| add learnable softmax temperatures per level (ablation) | ASpanFormer · ScanNet | pose AUC@20deg: 62.3 -> 63.3 | Improved |
| use 3x3 conv FFN in place of self attention plus MLP FFN (ablation) | ASpanFormer · ScanNet | pose AUC@5deg: 24.8 -> 25.6 | Improved |
| add normalized positional encoding when test resolution differs from training (ablation) | ASpanFormer · MegaDepth | pose AUC@5deg: 52.8 -> 55.3 | Improved |
| add normalized positional encoding (flow accuracy) (ablation) | ASpanFormer · MegaDepth | flow accuracy: 22.6 -> 72.3 | Improved |
| increase GLA iterations for flow regression (ablation) | ASpanFormer · ScanNet | flow accuracy <6px (%): 69.1 -> 72.3 | Improved |
| increase GLA iterations: uncertainty of unmatchable pixels (ablation) | ASpanFormer · ScanNet | mean 5 sigma (px) unmatchable: 19.4 -> 27.1 | Improved |
| replace linear attention (LoFTR) with ASpanFormer attention module: runtime | ASpanFormer vs LoFTR · ScanNet 640x480 | total runtime (ms): 97.7 -> 113.5 | Hurt |
| replace LoFTR with ASpanFormer in HLoc | HLoc · InLoc DUC2 | localization accuracy (1m,10deg): 85.5 -> 81.7 | Hurt |
| replace SuperGlue with ASpanFormer in HLoc | HLoc · InLoc DUC1 | localization accuracy (0.25m,2deg): 49.0 -> 51.5 | Improved |
Later papers that compared with this method
8 later accepted papers report ASpanFormer as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- Scalable Feature Matching via State Space Modeling and Sparse Correlation (CVPR 2026 · better on 1 of 2 shared results)
- SGAT: Learning Feature Matching with Singularity-enhanced Graph Attention Network (AAAI 2026 · better on 2 of 2 shared results)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · better on 4 of 4 shared results)
- CoMatch: Dynamic Covisibility-Aware Transformer for Bilateral Subpixel-Level Semi-Dense Image Matching (ICCV 2025 · better on 4 of 4 shared results)
- JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba (CVPR 2025 · better on 2 of 3 shared results)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · better on 1 of 2 shared results)
- ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses (NeurIPS 2024 · better on 0 of 1 shared result)
- Adaptive Spot-Guided Transformer for Consistent Local Feature Matching (CVPR 2023 · better on 5 of 6 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
- TransforMatcher: Match-to-Match Attention for Semantic Correspondence (CVPR 2022 · similarity 0.89)
- Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · similarity 0.87)
- BiFormer: Vision Transformer With Bi-Level Routing Attention (CVPR 2023 · similarity 0.87)
- CoMatch: Dynamic Covisibility-Aware Transformer for Bilateral Subpixel-Level Semi-Dense Image Matching (ICCV 2025 · similarity 0.87)
- Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence (ICLR 2024 · similarity 0.85)
- ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses (NeurIPS 2024 · similarity 0.85)
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.
