Research map / Graph learning / Graph self-supervised learning
Graph Context Transformation Learning for Progressive Correspondence Pruning
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 GCT-Net, a progressive two-view correspondence pruning network that builds multi-branch graph contexts with self- and cross-attention and uses confidence-based sampling with a transformer for global consensus guidance.
Problem
Outlier removal (correspondence pruning) for two-view matching where initial correspondences often have over 80% outliers; existing learning methods gather context but do not effectively use it.
Approach
Generate two complementary graph contexts (credible MLP/maxpool-based and structure convolution-based), refine each with self-attention and cross-attention, then use a Graph Context Guidance Transformer that samples high-confidence correspondences as guiding source to propagate global consensus to the rest.
Key results
- On YFCC100M classification, GCT-Net reaches 78.00% F-score versus 75.73% for CLNet (second best by 2.27%).
- On SUN3D classification, GCT-Net achieves 64.31% F-score, 0.51% above the second-best method.
- On YFCC100M with SIFT, GCT-Net gets 63.80 mAP@5° on unknown scenes, surpassing CLNet by about 13%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GCET to IPS (ablation) | GCT-Net · YFCC100M known, SIFT | mAP5°: +13.97% | Helped |
| Add GCET to IPS (ablation) | GCT-Net · YFCC100M known, SIFT | mAP20°: +10.27% | Helped |
| Add partial GCGT (sampling and consensus guidance) to IPS (ablation) | GCT-Net · YFCC100M known, SIFT | mAP5°: +11.67% | Helped |
| Add OAFilter to partial GCGT (ablation) | GCT-Net · YFCC100M known, SIFT | mAP5°: +1.3% | Helped |
| Add OAFilter to partial GCGT (ablation) | GCT-Net · YFCC100M known, SIFT | mAP20°: +0.84% | Helped |
| Add GCET and full GCGT to IPS (ablation) | GCT-Net · YFCC100M known, SIFT | mAP5°: 46.63 -> 63.80 | Helped |
| Add GCET and full GCGT to IPS (ablation) | GCT-Net · YFCC100M known, SIFT | mAP20°: 71.06 -> 82.86 | Helped |
| Use sampling rate 0.05 instead of the selected 0.2 (ablation) | GCT-Net GCGT · YFCC100M | mAP | Hurt |
| Use sampling rate 0.5 instead of the selected 0.2 (ablation) | GCT-Net GCGT · YFCC100M | mAP | Hurt |
| Use GCT-Net with SIFT features rather than ConvMatch | Relative pose estimation · YFCC100M known and unknown scenes | mAP5°: +9.18% and 6.98% improvements compared to ConvMatch on unknown and known scenes | Helped |
| Use GCT-Net with SuperPoint instead of ConvMatch | Relative pose estimation · YFCC100M known and unknown scenes | mAP5° | Hurt |
| Use GCT-Net with SuperPoint instead of ConvMatch | Relative pose estimation · YFCC100M known and unknown scenes | mAP20° | Helped |
Surprising findings
- With SuperPoint features, GCT-Net trails ConvMatch in mAP@5° and is only marginally better in mAP@20°, which the authors attribute to pruning already high-quality correspondences.
Limitations
- Progressive pruning discards hidden inliers, giving lower recall than several baselines.
- Performance gain over ConvMatch diminishes with SuperPoint features at mAP@5°.
Experimental setup
Datasets: YFCC100M SUN3D
Benchmarks: YFCC100M SUN3D
Metrics: precision recall F-score mAP@5° mAP@20°
Baselines: RANSAC LFGC OANet++ ACNet T-Net CLNet MS$^2$DG-Net MSA ConvMatch
Builds on: CLNet ConvMatch OANet LFGC-Net weighted eight-point algorithm
Closest papers by meaning
- BCLNet: Bilateral Consensus Learning for Two-View Correspondence Pruning (AAAI 2024 · similarity 0.86)
- MGNet: Learning Correspondences via Multiple Graphs (AAAI 2024 · similarity 0.85)
- VSFormer: Visual-Spatial Fusion Transformer for Correspondence Pruning (AAAI 2024 · similarity 0.85)
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.85)
- GPEN: Global Position Encoding Network for Enhanced Subgraph Representation Learning (ICML 2025 · similarity 0.84)
- A Scalable and Effective Alternative to Graph Transformers (AAAI 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.