Research map / 3D vision / 3D generation
CoMatcher: Multi-View Collaborative Feature Matching
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 CoMatcher, a deep multi-view sparse feature matcher with multi-view cross-attention, geometry-constrained attention, and confidence-guided correlation, plus a groupwise pipeline that scales it to large image sets.
Problem
Pairwise two-view feature matching yields ambiguous correspondences under severe occlusions, repetitive textures and extreme viewpoint changes, and errors compound when merged into multi-view tracks for SfM and localization.
Approach
Jointly match a target view against a group of co-visible source views in one network, using cross-view projection geometry from precomputed tracks to constrain attention and confidence-weighted attention correlation to enforce cross-view consistency; a co-visibility grouping step scales it to large collections.
Key results
- On HPatches homography (RANSAC), SP+CoMatcher reaches 38.4 / 68.9 / 79.0 AUC at 1/3/5px, above SP+LightGlue's 37.2 / 67.8 / 78.1.
- On MegaDepth relative pose (RANSAC), DISK+CoMatcher gets 54.9 / 71.2 / 82.0 AUC at 5/10/20 degrees versus 53.2 / 69.2 / 80.2 for DISK+LightGlue, at 73 ms per quadruplet.
- On IMC 2020, DISK+CoMatcher reaches 44.9 / 57.3 stereo and 67.1 / 78.4 multi-view AUC at 5/10 degrees, above DISK+LightGlue's 42.1 / 55.6 and 65.2 / 76.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-view collaborative matching instead of pairwise matching | Feature matching with DISK · IMC 2020 stereo | AUC @10°: 55.6 -> 57.3 | Helped |
| Use multi-view collaborative matching instead of pairwise matching | Feature matching with DISK · IMC 2020 multi-view reconstruction | AUC @10°: 76.2 -> 78.4 | Helped |
| Add multi-view feature interaction (ablation) | CoMatcher · Synthetic homography | precision: 89.7 -> 92.7 | Helped |
| Add geometric constraint mechanism (attention propagation) (ablation) | CoMatcher · Synthetic homography | precision: 87.7 -> 92.7 | Helped |
| Add multi-view feature correlation (ablation) | CoMatcher · Synthetic homography | precision: 90.5 -> 92.7 | Helped |
| Increase the source-view group size (ablation) | CoMatcher · Synthetic homography | AUC | Mixed |
| Use a N-to-1 architecture instead of N-to-N multi-view matching | CoMatcher · HPatches | AUC-DLT: 34.3 / 66.9 / 75.5 -> 37.1 / 69.0 / 78.8 | Helped |
Surprising findings
- End2End, another joint multi-view matcher, performs notably worse than CoMatcher and sometimes below pairwise LightGlue.
- Increasing the number of source views beyond about 5 degrades matching performance on the synthetic homography dataset.
Limitations
- Scalability is limited by the search space and redundant information when many source views are used; performance declines beyond about 5 views.
- Groupwise pipeline depends on precomputed tracks from existing frameworks for the group-level geometry cues.
Experimental setup
Datasets: synthetic homography dataset MegaDepth HPatches IMC 2020 phototourism
Benchmarks: HPatches MegaDepth IMC 2020
Metrics: AUC at 1px/3px/5px AUC at 5°/10°/20° AUC 5°/10° for stereo and multi-view precision recall runtime
Baselines: SIFT+NN+mutual SuperPoint+NN+mutual SuperGlue SGMNet LightGlue DISK+LightGlue End2End LoFTR PDC-Net DUSt3R MASt3R
Builds on: LightGlue SuperGlue End2End LoFTR
Closest papers by meaning
- MATCHA: Towards Matching Anything (CVPR 2025 · similarity 0.86)
- SegMASt3R: Geometry Grounded Segment Matching (NeurIPS 2025 · similarity 0.86)
- ADA-Track: End-to-End Multi-Camera 3D Multi-Object Tracking with Alternating Detection and Association (CVPR 2024 · similarity 0.85)
- Bridging Viewpoint Gaps: Geometric Reasoning Boosts Semantic Correspondence (CVPR 2025 · similarity 0.85)
- MITracker: Multi-View Integration for Visual Object Tracking (CVPR 2025 · similarity 0.84)
- Learning Intra-view and Cross-view Geometric Knowledge for Stereo Matching (CVPR 2024 · similarity 0.84)
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.