Research map / Graph learning / Graph algorithms and clustering
FC-GNN: Recovering Reliable and Accurate Correspondences from Interferences
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 FC-GNN, a k-nearest-neighbor attention graph network that jointly filters outlier correspondences and calibrates inlier pixel offsets using match geometry and patch appearance.
Problem
Sparse local feature matching is degraded by matching noise (outliers) and inaccurate keypoint localization, and prior filtering-only or refinement methods either cannot fix individual match quality or are local and tied to dense pipelines.
Approach
Treat correspondence optimization as joint filtering (confidence scores that reject outliers) and calibration (per-match pixel offsets), using a graph attention network over matches where each match attends only to its k nearest matches in matching space, combined with embedded local patch pairs.
Key results
- On HPatches, FC-GNN raises SP + SG mean matching accuracy at 1 px from about 40% to 67%.
- On MegaDepth1500 outdoor pose estimation, FC-GNN improves SP + SG AUC@5° from 49.13 to 54.67.
- FC-GNN runs in about 6.4 ms on HPatches with SP + SG, versus 104.4 ms for Patch2Pix refinement with the same matcher.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FC-GNN filtering and calibration | SP + SuperGlue · HPatches overall | Homography AUC @3px (%): 57.0 -> 67.0 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · HPatches overall | Homography AUC @5px (%): 69.2 -> 77.2 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · HPatches overall | Homography AUC @10px (%): 81.7 -> 86.9 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · MegaDepth1500 outdoor | Pose estimation AUC @5°: 47.32 -> 54.67 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · MegaDepth1500 outdoor | Pose estimation AUC @10°: 63.98 -> 71.03 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · MegaDepth1500 outdoor | Pose estimation AUC @20°: 77.23 -> 82.65 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · ScanNet indoor | Pose estimation AUC @5°: 15.33 -> 18.46 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · ScanNet indoor | Pose estimation AUC @10°: 31.74 -> 36.47 | Helped |
| add FC-GNN filtering and calibration | SP + SuperGlue · ScanNet indoor | Pose estimation AUC @20°: 48.10 -> 52.98 | Helped |
| add both filtering and calibration instead of filtering alone (ablation) | FC-GNN with context information (T) · HPatches | MMA @1px: 39.9 -> 52.3 | Helped |
| add both filtering and calibration instead of calibration alone (ablation) | FC-GNN with context information (T) · HPatches | MMA @1px: 43.3 -> 52.3 | Helped |
| add both filtering and calibration instead of filtering alone (ablation) | FC-GNN with local patch information (L) · HPatches | MMA @1px: 45.3 -> 59.6 | Helped |
Surprising findings
- OANet and Patch2Pix degrade SP + SG pose estimation (e.g., OANet AUC@5° drops from 49.13 to 23.40), indicating that existing refiners can hurt strong matchers.
Limitations
- Training uses synthetic noise on MegaDepth rather than real matcher outputs, which may not match real error distributions.
- Dense matcher gains are smaller (ASpanFormer improves less than sparse matchers), and OANet/Patch2Pix hurt ASpanFormer.
Experimental setup
Datasets: MegaDepth HPatches MegaDepth1500 ScanNet
Benchmarks: HPatches MegaDepth1500 ScanNet
Metrics: mean matching accuracy homography AUC at 3/5/10 px pose estimation AUC at 5/10/20 degrees runtime
Baselines: OANet Patch2Pix Origin (raw matching) ASpanFormer
Builds on: SuperGlue OANet Patch2Pix Transformer self-attention
Closest papers by meaning
- GIM: Learning Generalizable Image Matcher From Internet Videos (ICLR 2024 · similarity 0.86)
- Improving GNN Calibration with Discriminative Ability: Insights and Strategies (AAAI 2024 · similarity 0.86)
- MaxEnt Loss: Calibrating Graph Neural Networks under Out-of-Distribution Shift (Student Abstract) (AAAI 2024 · similarity 0.85)
- FastJAM: a Fast Joint Alignment Model for Images (NeurIPS 2025 · similarity 0.85)
- OmniGlue: Generalizable Feature Matching with Foundation Model Guidance (CVPR 2024 · similarity 0.85)
- Learning to Filter Outlier Edges in Global SfM (CVPR 2025 · 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.