Research map / Graph learning / Knowledge and temporal graphs
MGNet: Learning Correspondences via Multiple Graphs
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 MGNet, which combines implicit (DiffPooling-based) and explicit (kNN graph) local graphs into a global graph via Graph Soft Degree Attention to reject outliers in sparse correspondences.
Problem
Finding correct correspondences (inliers) from uneven, low-inlier-rate putative correspondence sets, where prior GNN methods build a single graph type or naively stack local graphs.
Approach
Construct implicit and explicit local graphs in sequence and combine them into a global graph, using Graph Soft Degree Attention (a soft-adjacency-based degree matrix Hadamard-multiplied with node features, with no extra parameters) to capture and amplify discriminative features.
Key results
- On YFCC100M outdoor with SIFT, MGNet reaches 64.63 mAP@5° on unknown scenes versus 60.53 for U-Match (second best).
- On HPatches, MGNet reaches 52.08% accuracy at 3px versus 48.90% for U-Match.
- Ablation: GSDA global edge construction achieves 64.63 mAP@5° vs 63.08 for plain GCN, with MGNet at 1.31 MB model size.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MGNet instead of U-Match | Camera pose estimation with SIFT · YFCC100M/SUN3D unknown outdoor | mAP5° (%): 60.53 -> 64.63 | Helped |
| Use MGNet instead of U-Match | Camera pose estimation with SIFT · YFCC100M/SUN3D unknown indoor | mAP5° (%): 21.38 -> 21.27 | Hurt |
| Use GSDA instead of a plain GCN for global edges (ablation) | MGNet · Unknown outdoor scene with SIFT | mAP5°: 63.08 -> 64.63 | Helped |
| Use GSDA instead of a plain GCN for global edges (ablation) | MGNet · Unknown outdoor scene with SIFT | mAP20°: 82.72 -> 83.76 | Helped |
| Use max-pooling and MLPs instead of average-pooling and MLPs (ablation) | Explicit local graph aggregation in MGNet · Unknown outdoor scene with SIFT | mAP5°: 58.35 -> 61.35 | Helped |
| Add implicit local graph before explicit local graph (ablation) | MGNet local graph construction · Unknown outdoor scene with SIFT | mAP5°: 59.95 -> 61.35 | Helped |
| Avoid adding implicit and explicit local probabilities to verification (ablation) | MGNet · Unknown outdoor scene with SIFT | mAP5°: 61.23 -> 64.63 | Helped |
| Add the verification framework (ablation) | MGNet without verification · Unknown outdoor scene with SIFT | mAP5°: 61.17 -> 64.63 | Helped |
| Use pruning with ratio 0.75 (ablation) | MGNet verification framework · Unknown outdoor scene with SIFT | mAP5°: 64.63 -> 50.80 | Hurt |
| Use pruning with ratio 0.5 (ablation) | MGNet verification framework · Unknown outdoor scene with SIFT | mAP5°: 64.63 -> 61.68 | Hurt |
| Use pruning with ratio 0.25 (ablation) | MGNet verification framework · Unknown outdoor scene with SIFT | mAP5°: 64.63 -> 63.05 | Hurt |
| Add implicit local graph while retaining verification (ablation) | Verification-only model · Unknown outdoor scene with SIFT | mAP5°: 44.9 -> 55.2 | Helped |
Surprising findings
- The pruning operation hurts performance; at pruning ratio 0.25 the model performs worse than without the verification framework.
- Adding extra local (implicit or explicit) probabilities to the global probability generally degrades performance, attributed to overfitting.
Limitations
- Gains are modest on some indoor metrics (e.g., unknown indoor mAP@5° 21.27 vs 21.38 for U-Match)
- Ablations rely on a single outdoor unknown scene setting
Experimental setup
Datasets: YFCC100M SUN3D
Benchmarks: YFCC100M SUN3D PhotoTourism HPatches Aachen Day-Night
Metrics: mAP@5° mAP@20° homography accuracy at 3/5/10 px localization accuracy at distance/orientation thresholds
Baselines: RANSAC PointNet++ DFE CNe OA-Net++ ACNe SuperGlue LMC-Net CL-Net MS2DG-Net U-Match
Builds on: OA-Net CL-Net LMC-Net DiffPooling OA-Net++
Closest papers by meaning
- Graph Context Transformation Learning for Progressive Correspondence Pruning (AAAI 2024 · similarity 0.85)
- Sketch-Augmented Features Improve Learning Long-Range Dependencies in Graph Neural Networks (NeurIPS 2025 · similarity 0.85)
- Iteratively Refined Early Interaction Alignment for Subgraph Matching based Graph Retrieval (NeurIPS 2024 · similarity 0.84)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.84)
- Unified Graph Neural Networks Pre-training for Multi-domain Graphs (AAAI 2025 · similarity 0.84)
- Multimodal Graph Neural Architecture Search under Distribution Shifts (AAAI 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.