Research map / 3D vision / Human pose estimation
Learning Affine Correspondences by Integrating Geometric Constraints
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 a pipeline that extracts affine correspondences by pairing a DKM-based dense matcher trained with a Sampson-distance epipolar loss and a local affine shape extractor trained with an affine Sampson loss.
Problem
Extracting accurate affine correspondences in real-world scenes, where detector-based methods struggle with repetitive or weak textures and view-synthesis methods are computationally expensive.
Approach
Replace sparse detectors with dense matching to get point correspondences, then estimate local affine shape (orientation, scale, residual shape) with a network supervised by novel epipolar-based Sampson distance losses (PC-loss and affine Sampson loss).
Key results
- Achieves overall HPatches MMAscore of 0.851, above RoMA's 0.843.
- On MegaDepth-1500 with affine GC-RANSAC, reaches AUC@5° of 65.5 versus 60.4 for DKM and 62.6 for RoMA.
- On KITTI, average rotation RMSE 0.0289° versus 0.0343° for RoMA; affine shape cosine similarity 0.994 vs 0.894 for ASIFT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed affine correspondence method instead of RoMA | Image matching · HPatches overall | MMAscore: 0.843 -> 0.851 | Helped |
| Use the proposed affine correspondence method instead of RoMA | Image matching · HPatches illumination | MMAscore: 0.901 -> 0.908 | Helped |
| Use the proposed affine correspondence method instead of RoMA | Image matching · HPatches viewpoint | MMAscore: 0.789 -> 0.798 | Helped |
| Use affine GC-RANSAC with the proposed affine correspondences instead of RoMA | Relative pose estimation · MegaDepth-1500 | AUC @ 5°: 62.6 -> 65.5 | Helped |
| Use affine GC-RANSAC with the proposed affine correspondences instead of RoMA | Relative pose estimation · MegaDepth-1500 | AUC @ 10°: 76.7 -> 77.9 | Helped |
| Use affine GC-RANSAC with the proposed affine correspondences instead of RoMA | Relative pose estimation · MegaDepth-1500 | AUC @ 20°: 86.3 -> 86.3 | No clear effect |
| Add PC-loss to DKM (ablation) | DKM · MegaDepth-1500 | AUC @ 5°: 60.4 -> 60.7 | Helped |
| Add PC-loss to DKM (ablation) | DKM · MegaDepth-1500 | AUC @ 10°: 74.9 -> 75.1 | Helped |
| Add PC-loss to DKM (ablation) | DKM · MegaDepth-1500 | AUC @ 20°: 85.1 -> 85.4 | Helped |
| Add the affine extractor and AC-loss to DKM + PC-loss (ablation) | DKM + PC-loss · MegaDepth-1500 | AUC @ 5°: 60.7 -> 63.1 | Helped |
| Add the affine extractor and AC-loss to DKM + PC-loss (ablation) | DKM + PC-loss · MegaDepth-1500 | AUC @ 10°: 75.1 -> 76.3 | Helped |
| Add the affine extractor and AC-loss to DKM + PC-loss (ablation) | DKM + PC-loss · MegaDepth-1500 | AUC @ 20°: 85.4 -> 85.2 | Hurt |
Surprising findings
- Plain RANSAC with the proposed matcher gives only similar AUC to RoMA; gains mainly appear when affine shapes are used in GC-RANSAC.
Limitations
- Affine shape extractor relies on a frozen dense matcher trained separately, so errors in point matches propagate.
- Training is decoupled with weak supervision, and joint training was not reported as a fair comparison.
Experimental setup
Datasets: MegaDepth ScanNet HPatches KITTI
Benchmarks: HPatches MegaDepth-1500 KITTI
Metrics: MMA score Euclidean distance cosine similarity AUC at 5/10/20 degrees rotation RMSE translation RMSE
Baselines: DKM RoMA LoFTR ASpanFormer PDC-Net+ AffNet ASIFT VLFeat DISK PoSFeat
Builds on: DKM AffNet GC-RANSAC PoSFeat
Closest papers by meaning
- From Correspondences to Pose: Non-minimal Certifiably Optimal Relative Pose without Disambiguation (CVPR 2024 · similarity 0.87)
- Structure-Aware Correspondence Learning for Relative Pose Estimation (CVPR 2025 · similarity 0.86)
- Rethinking Correspondence-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.86)
- Relative Pose Estimation through Affine Corrections of Monocular Depth Priors (CVPR 2025 · similarity 0.85)
- HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation (CVPR 2024 · similarity 0.85)
- FAR: Flexible Accurate and Robust 6DoF Relative Camera Pose Estimation (CVPR 2024 · similarity 0.85)
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.