Research map / 3D vision / Depth and stereo
BCLNet: Bilateral Consensus Learning for Two-View 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 BCLNet, a two-view correspondence pruning network that runs a global self-attention block (BCMA) in parallel with local Order-Aware context learning, fuses them into bilateral consensus, and recalibrates with a BCR block.
Problem
Two-view correspondence pruning (classifying inlier correspondences and recovering camera pose) where prior methods process local and global context progressively, neglecting their interaction.
Approach
Parallel local and global consensus learning: a channel-wise linear-complexity attention with KNN-based local embedding (BCMA) runs alongside the Order-Aware block; their concatenation is refined by a bilateral consensus recalibration (BCR) block combining global SE-style pooling and local KNN features.
Key results
- On YFCC100M with SIFT, BCLNet reaches 66.08% mAP5° on unknown scenes, a 3.98% gain over the second-best method (NCMNet at 62.10%).
- On YFCC100M known scenes, BCLNet achieves 52.62% mAP5° vs 49.10% for NCMNet with SIFT.
- Ablation on YFCC100M unknown scenes: adding BCMA raises mAP5° from 51.80 to 59.98; adding BCR and OA brings it to 66.08.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BCLNet instead of NCMNet | Correspondence classification · YFCC100M known | F-score (%): 79.69 -> 80.10 | Helped |
| Use BCLNet instead of NCMNet | Correspondence classification · YFCC100M unknown | F-score (%): 77.41 -> 78.31 | Helped |
| Use BCLNet instead of NCMNet | Camera pose estimation with SIFT · YFCC100M known | mAP5° (%): 49.10 -> 52.62 | Helped |
| Use BCLNet instead of NCMNet | Camera pose estimation with SIFT · YFCC100M unknown | mAP5° (%): 62.10 -> 66.08 | Helped |
| Use BCLNet instead of ConvMatch | Camera pose estimation with SuperPoint · YFCC100M unknown | mAP5° (%): 48.80 -> 48.07 | Hurt |
| Add BCMA to the iterative pruning strategy (ablation) | BCLNet · YFCC100M unknown | mAP5° (%): 51.80 -> 59.98 | Helped |
| Add BCMA to the iterative pruning strategy (ablation) | BCLNet · YFCC100M unknown | mAP20° (%): 75.76 -> 80.72 | Helped |
| Add BCR to IPS + BCMA (ablation) | BCLNet · YFCC100M unknown | mAP5° (%): 59.98 -> 61.03 | Helped |
| Add BCR to IPS + BCMA (ablation) | BCLNet · YFCC100M unknown | mAP20° (%): 80.72 -> 81.18 | Helped |
| Add OA to IPS + BCMA + BCR (ablation) | BCLNet · YFCC100M unknown | mAP5° (%): 61.03 -> 66.08 | Helped |
| Add OA to IPS + BCMA + BCR (ablation) | BCLNet · YFCC100M unknown | mAP20° (%): 81.18 -> 83.38 | Helped |
Surprising findings
- On SUN3D, BCLNet is slightly behind ConvMatch (e.g., 19.96 vs 21.71 mAP5° on unknown scenes) despite gains on outdoor data.
Limitations
- Slightly lower than ConvMatch on indoor SUN3D benchmarks.
- Recall drops relative to some baselines due to pruning, which removes inliers.
Experimental setup
Datasets: YFCC100M SUN3D
Benchmarks: YFCC100M SUN3D
Metrics: precision recall F-score mAP5° mAP20°
Baselines: RANSAC PointNet++ LFGC OA-Net++ MSA-Net CLNet MS2DG-Net ConvMatch NCMNet
Builds on: CLNet OANet NCMNet SENet LFGC
Closest papers by meaning
- DeMo: Deep Motion Field Consensus with Learnable Kernels for Two-view Correspondence Learning (AAAI 2025 · similarity 0.87)
- Graph Context Transformation Learning for Progressive Correspondence Pruning (AAAI 2024 · similarity 0.86)
- Self-Supervised Cross-View Correspondence with Predictive Cycle Consistency (CVPR 2025 · similarity 0.84)
- Spotting the Unseen: Reciprocal Consensus Network Guided by Visual Archetypes (AAAI 2024 · similarity 0.84)
- DeMatch: Deep Decomposition of Motion Field for Two-View Correspondence Learning (CVPR 2024 · similarity 0.84)
- VSFormer: Visual-Spatial Fusion Transformer for Correspondence Pruning (AAAI 2024 · 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.