Research map / Architectures and efficiency / Attention and transformers
ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses
Read the paperCheck your idea against 108,967 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 ETO, an efficient transformer-based local feature matcher that organizes multiple homography hypotheses to reduce transformer tokens in coarse matching and uses one-way cross-attention in refinement, achieving several times faster inference than transformer-based baselines.
Problem
Transformer-based local feature matching is accurate but too slow for low-latency or large-scale applications because attention cost grows quadratically with the number of image tokens.
Approach
Coarse matches are organized as multiple local homography hypotheses, each covering several patches, so far fewer tokens (1/32 resolution) enter the transformer; a segmentation step re-selects the best hypothesis per sub-unit, and refinement replaces bidirectional self/cross-attention with a single uni-directional cross-attention.
Key results
- On MegaDepth outdoor pose estimation, ETO reaches 21.0 ms runtime with AUC@5 of 51.7, versus LoFTR at 93.2 ms with AUC@5 of 57.5.
- On YFCC100M, ETO achieves AUC@5/10/20 of 44.8/64.0/78.8 at 22.1 ms, comparable to LoFTR (44.7/63.6/78.3 at 96.3 ms).
- On ScanNet, ETO attains AUC@20 of 59.8 at 24.2 ms, close to LoFTR's 57.2 at 94.2 ms (note: AUC@5 of 20.1 is well below LoFTR's 21.4 at 94.2 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Base32 matching without homography with Base32 matching using homography (ablation) | ETO coarse matching · MegaDepth | pose AUC @5°: 9.5 -> 28.5 | Helped |
| replace Base8 matching without homography with Base32 matching using homography (ablation) | ETO coarse matching · MegaDepth | time (ms): 58.9 -> 8.5 | Helped |
| use Base32 matching with homography instead of Base8 matching without homography (ablation) | ETO coarse matching · MegaDepth | pose AUC @5°: 25.7 -> 28.5 | Helped |
| use unidirectional attention instead of full attention for refinement (ablation) | ETO refinement · MegaDepth | time (ms): 32.8 -> 22.0 | Helped |
| use unidirectional attention instead of full attention for refinement (ablation) | ETO refinement · MegaDepth | pose AUC @5°: 51.0 -> 51.7 | No clear effect |
| add segmentation to the unidirectional refinement (ablation) | ETO · MegaDepth | pose AUC @5°: 42.1 -> 51.7 | Helped |
| add homography hypotheses to Base32 matching (ablation) | ETO coarse matching · MegaDepth | pose AUC @10°: 21.3 -> 46.2 | Helped |
| add segmentation to unidirectional refinement (ablation) | ETO · MegaDepth | time (ms): 21.2 -> 22.0 | Hurt |
Surprising findings
- On MegaDepth, ETO's AUC@5 (51.7) trails LoFTR (57.5) and ASpanFormer (58.6), so its speed gain comes with an accuracy cost at strict thresholds.
- Unidirectional cross-attention in refinement is reported to match the accuracy of full bidirectional attention while being substantially faster.
Limitations
- Matching accuracy remains below state-of-the-art dense/detector-free methods such as RoMa and DKM, as the authors acknowledge.
- The parametric homography scheme (rotation, scale, perspective) may not be optimal or complete for representing local transformations.
- The feature extractor is not trained end-to-end, limiting fine-grained enhancement.
Experimental setup
Datasets: MegaDepth ScanNet
Benchmarks: HPatches MegaDepth YFCC100M ScanNet-1500
Metrics: AUC@5 AUC@10 AUC@20 average corner error point accuracy at 1px runtime (ms)
Baselines: LoFTR ASpanFormer Quadtree LightGlue SuperGlue SuperPoint+search RoMa Tiny-RoMa Efficient LoFTR
Builds on: LoFTR LightGlue SuperGlue SuperPoint
Closest papers by meaning
- JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba (CVPR 2025 · similarity 0.87)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.86)
- MATCHA: Towards Matching Anything (CVPR 2025 · similarity 0.86)
- Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence (ICLR 2024 · similarity 0.86)
- RoMa: Robust Dense Feature Matching (CVPR 2024 · similarity 0.85)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · 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.