Research map / 3D vision / Human pose estimation
Learning Feature Descriptors using Camera Pose Supervision
Qianqian Wang, Xiaowei Zhou, Bharath Hariharan, Noah Snavely
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CAPS descriptors with SuperPoint keypoints instead of SuperPoint descriptors | SuperPoint keypoints · HPatches homography estimation (eps=1px) | corner correctness accuracy (%): 37.4 -> 44.8 | Improved |
| use CAPS descriptors with SIFT keypoints instead of SIFT descriptors | SIFT keypoints · HPatches homography estimation (eps=3px) | corner correctness accuracy (%): 68.1 -> 72.2 | Improved |
| use CAPS descriptors with SIFT keypoints instead of SIFT descriptors | SIFT keypoints · HPatches homography estimation (eps=1px) | corner correctness accuracy (%): 40.5 -> 34.6 | Hurt |
| use CAPS with SuperPoint keypoints instead of SuperPoint descriptors | SuperPoint keypoints · MegaDepth relative pose (hard) | rotation accuracy (%): 24.5 -> 38.1 | Improved |
| use CAPS with SIFT keypoints instead of SIFT with ratio test | SIFT keypoints · MegaDepth relative pose (hard) | rotation accuracy (%): 20.8 -> 36.8 | Improved |
| use CAPS with SuperPoint keypoints instead of R2D2 | R2D2 · ScanNet relative pose (dframe=60) | rotation accuracy (%): 62.9 -> 59.3 | Hurt |
| use CAPS instead of GeoDesc/SOSNet descriptors | SOSNet · ETH benchmark Madrid Metropolis (# registered images) | registered images: 844 -> 851 | No clear effect |
| train with camera-pose epipolar loss instead of ground-truth correspondence triplet loss (ablation) | CAPS · HPatches ablation (MMA, PCK) | MMA and PCK | Improved |
| train with ground-truth correspondences (L2 loss) instead of camera pose only (ablation) | CAPS · HPatches ablation (MMA, PCK) | MMA and PCK | Mixed |
| add cycle consistency loss (ablation) | CAPS · HPatches ablation (MMA, PCK) | MMA and PCK | Improved |
| train with cycle consistency loss only, no epipolar loss (ablation) | CAPS · HPatches ablation (MMA, PCK) | MMA and PCK | Hurt |
| add coarse-to-fine architecture (ablation) | CAPS · HPatches ablation (MMA, PCK) | MMA and PCK | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GeoDesc — GeoDesc: Learning Local Descriptors by Integrating Geometry Constraints (ECCV 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Decoupling Makes Weakly Supervised Local Feature Better (CVPR 2022 · similarity 0.87)
- Self-Supervised Learning of Interpretable Keypoints From Unlabelled Videos (CVPR 2020 · similarity 0.87)
- End-to-End Learning of Keypoint Detector and Descriptor for Pose Invariant 3D Matching (CVPR 2018 · similarity 0.87)
- Unsupervised Learning of Local Equivariant Descriptors for Point Clouds (TPAMI 2021 · similarity 0.86)
- Self-Supervised Geometric Perception (CVPR 2021 · similarity 0.86)
- Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.86)
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.
