Research map / 3D vision / Human pose estimation
Dense Match Summarization for Faster Two-view 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.
Proposes a clustering-based match summarization scheme that replaces dense correspondences with ~1% representative matches plus 9x9 proxy matrices, speeding up RANSAC two-view relative pose estimation 10-100x with comparable accuracy.
Problem
Robust two-view relative pose estimation from dense correspondences is accurate but RANSAC scoring and refinement become very slow because of the large number of matches.
Approach
Cluster dense matches by 4D position (K-means), keep one representative match per cluster for RANSAC, and summarize each cluster's Sampson residuals into a single 9x9 reduced measurement matrix so refinement approximates the full dense cost at low runtime.
Key results
- CCC variant with DKM on MegaDepth-1500 achieves AUC@5° 66.0 vs 67.4 for dense baseline at 55.0x speedup (1.2 ms vs 66.0 ms).
- CCA variant with RoMa on MegaDepth-1500 matches dense AUC@5° (70.1) at 42.2x speedup.
- On WxBS with RoMa, CCA reaches recall 85.1 vs 86.3 dense at 43.3x speedup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Summarize RoMa matches with representative matches and approximate residuals instead of using all dense matches | RoMa two-view estimation · MegaDepth-1500 | RT (ms): 60.6 -> 1.4 | Helped |
| Summarize RoMa matches with representative matches and approximate residuals instead of using all dense matches | RoMa two-view estimation · MegaDepth-1500 | AUC@5°: 70.1 -> 70.1 | No clear effect |
| Summarize RoMa matches with representative matches and approximate residuals instead of using all dense matches | RoMa two-view estimation · ScanNet-1500 | RT (ms): 64.6 -> 1.4 | Helped |
| Summarize RoMa matches with representative matches and approximate residuals instead of using all dense matches | RoMa two-view estimation · ScanNet-1500 | AUC@5°: 33.3 -> 33.0 | No clear effect |
| Summarize RoMa matches with representative matches and approximate residuals instead of using all dense matches | RoMa fundamental matrix estimation · WxBS | RT (ms): 63.0 -> 1.45 | Helped |
| Summarize RoMa matches with representative matches and approximate residuals instead of using all dense matches | RoMa fundamental matrix estimation · WxBS | Recall (10px): 86.3 -> 85.1 | No clear effect |
| Add approximate residuals for refinement instead of refining with representative-match residuals (ablation) | DKM RANSAC (CCA vs CCC) · MegaDepth-1500 | AUC@5°: 65.95±0.10 -> 66.70±0.06 | Helped |
| Add approximate residuals for refinement instead of refining with representative-match residuals (ablation) | DKM RANSAC (CCA vs CCC) · MegaDepth-1500 | RT (ms): 1.2 -> 1.5 | Hurt |
| Use approximate residuals for scoring and dense matches for refinement instead of representative-match scoring (ablation) | DKM RANSAC (CAD vs DDD) · MegaDepth-1500 | AUC@5°: 67.38±0.10 -> 67.21±0.08 | No clear effect |
| Use representative-match residuals for scoring and dense matches for refinement instead of dense scoring (ablation) | DKM RANSAC (CCD vs DDD) · MegaDepth-1500 | RT (ms): 66.0 -> 5.3 | Helped |
| Use approximate residuals for refinement instead of dense-match refinement (ablation) | DKM RANSAC (CAA vs CAD) · MegaDepth-1500 | AUC@5°: 67.21±0.08 -> 66.87±0.07 | Hurt |
| Use approximate residuals for refinement instead of dense-match refinement (ablation) | DKM RANSAC (CAA vs CAD) · MegaDepth-1500 | RT (ms): 6.8 -> 2.7 | Helped |
Surprising findings
- Sparsified dense matches (K-means clustering) outperform state-of-the-art sparse SuperPoint+LightGlue matches in accuracy in the reported experiments.
- Subsampling the dense matcher output reduces accuracy more than the proposed summarization at equal runtime.
Limitations
- Only useful when the original number of matches is very large.
- Speeds up only the robust estimation step; the dense matcher itself dominates total runtime.
- Single-correspondence estimation from summarized matches generally performs worse.
Experimental setup
Datasets: MegaDepth-1500 ScanNet-1500 WxBS
Benchmarks: MegaDepth-1500 ScanNet-1500 WxBS
Metrics: AUC@5° AUC@10° AUC@20° runtime (ms) recall@10px
Baselines: DKM dense baseline RoMA dense baseline ASpanFormer dense baseline SuperPoint+LightGlue Grid clustering SLIC clustering subsampled DKM matches
Builds on: RANSAC LO-RANSAC PoseLib DKM MSAC
Closest papers by meaning
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · similarity 0.87)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.85)
- UFM: A Simple Path towards Unified Dense Correspondence with Flow (NeurIPS 2025 · similarity 0.85)
- RUBIK: A Structured Benchmark for Image Matching across Geometric Challenges (CVPR 2025 · similarity 0.84)
- MATCHA: Towards Matching Anything (CVPR 2025 · similarity 0.83)
- CoMatcher: Multi-View Collaborative Feature Matching (CVPR 2025 · 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.