Research map / Architectures and efficiency / RNNs and normalization
HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation
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.
Replaces the point-to-patch fine-matching step of semi-dense coarse-to-fine matchers (LoFTR, ASpanFormer) with an iterative patch-to-patch homography estimation module that yields sub-pixel and densifiable matches at semi-dense cost.
Problem
Fine-matching in semi-dense feature matching relies on point-to-patch expectation or direct regression, which is sensitive to irrelevant regions, lacks sub-pixel precision, and yields only one match per patch, hindering keypoint continuity and dense output.
Approach
Align coarse-matched patch pairs with an iteratively refined homography (parameterized by four corner displacements, updated via DLT) that samples a 4D correlation volume, so the whole patch is aligned and any point within it can be mapped, enabling cheap dense matching; the module is plug-in for existing semi-dense matchers.
Key results
- ASpan Homo improves AUC@5 on MegaDepth from 55.3 to 57.1 (+1.8) and on ScanNet from 19.6 to 22.0 (+2.4) over ASpanFormer.
- On MegaDepth EPE, ASpan Homo reaches PCK@1px 60.2 vs 54.3 for ASpanFormer and 63.7 for RoMa, at 442 ms vs RoMa's 1527 ms on a V100.
- The lightweight 5/1/1 configuration of the module runs in 13.7 ms versus 29.4 ms for the original ASpanFormer fine-matching module while raising PCK@1px from 54.3 to 55.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace LoFTR fine-matching with homography estimation | LoFTR · MegaDepth | AUC@5: 52.8 -> 55.1 | Helped |
| replace ASpanFormer fine-matching with homography estimation | ASpanFormer · MegaDepth | AUC@5: 55.3 -> 57.1 | Helped |
| replace ASpanFormer fine-matching with homography estimation | ASpanFormer · ScanNet | AUC@5: 19.6 -> 22.0 | Helped |
| replace ASpanFormer fine-matching with homography estimation | ASpanFormer · MegaDepth | PCK@1px: 54.3 -> 60.2 | Helped |
| replace ASpanFormer fine-matching with homography estimation | ASpanFormer · HPatches | AUC@3px: 67.4 -> 70.2 | Helped |
| add dense fine-level loss instead of semi-dense fine loss (ablation) | ASpan Homo · MegaDepth | PCK@1px: 58.6 -> 59.1 | Helped |
| add suppressed correct matches to fine-level supervision (ablation) | ASpan Homo · MegaDepth | PCK@1px: 59.1 -> 60.2 | Helped |
| increase patch size from w=5 to w=9, with r=1 and K=1 (ablation) | ASpan Homo · MegaDepth | PCK@1px: 55.6 -> 58.5 | Helped |
| increase homography iterations from K=1 to K=3, with w=9 and r=1 (ablation) | ASpan Homo · MegaDepth | PCK@1px: 58.5 -> 60.2 | Helped |
| increase search radius from r=1 to r=3, with w=9 and K=1 (ablation) | ASpan Homo · MegaDepth | PCK@1px: 58.5 -> 59.9 | Helped |
| increase homography iterations from K=1 to K=3, with w=9 and r=3 (ablation) | ASpan Homo · MegaDepth | PCK@1px: 59.9 -> 60.3 | Helped |
Surprising findings
- The lightweight 5/1/1 homography configuration is faster than the original fine-matching module while being more accurate.
Limitations
- Dense matches are only obtained within patches around coarse matches, so densification depends on coarse matching quality and re value.
- Heavier configurations increase runtime (e.g., 697 ms for the heavy version).
Experimental setup
Datasets: MegaDepth
Benchmarks: MegaDepth ScanNet HPatches
Metrics: AUC@5 AUC@10 AUC@20 PCK@1px PCK@3px PCK@5px runtime
Baselines: LoFTR ASpanFormer LoFTR (E) Affine-based Matcher SP+SG DKM RoMa SuperGlue
Builds on: LoFTR ASpanFormer IHN SuperGlue
Closest papers by meaning
- Adapting Dense Matching for Homography Estimation with Grid-based Acceleration (CVPR 2025 · similarity 0.90)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.88)
- UFM: A Simple Path towards Unified Dense Correspondence with Flow (NeurIPS 2025 · similarity 0.87)
- Noisy One-point Homographies are Surprisingly Good (CVPR 2024 · similarity 0.87)
- Dense Match Summarization for Faster Two-view Estimation (CVPR 2025 · similarity 0.87)
- RUBIK: A Structured Benchmark for Image Matching across Geometric Challenges (CVPR 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.