Research map / 3D vision / Human pose estimation
Learning to Make Keypoints Sub-Pixel Accurate
Shinjeong Kim, Marc Pollefeys, Daniel Barath
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 |
|---|---|---|---|
| add the learned sub-pixel refinement module to SuperPoint with mutual nearest neighbor matching | SuperPoint + MNN · MegaDepth relative pose | AUC@5: 35.34 -> 37.16 | Improved |
| add the refinement module to SuperPoint with LightGlue | SuperPoint + LightGlue · MegaDepth relative pose | AUC@5: 59.82 -> 61.82 | Improved |
| add the refinement module to ALIKED with LightGlue | ALIKED + LightGlue · MegaDepth relative pose | AUC@5: 64.69 -> 65.34 | Improved |
| use the proposed refinement instead of RL-based refinement (Reinforced SP) | SuperPoint + MNN · MegaDepth relative pose | AUC@5: 36.61 -> 37.16 | Improved |
| add the refinement module to SuperPoint + MNN on fundamental matrix estimation | SuperPoint + MNN · KITTI fundamental matrix | F1-score: 74.15 -> 75.48 | Improved |
| add SoftArgMax alone on top of SuperPoint without a CNN (ablation) | SuperPoint + MNN · MegaDepth relative pose | AUC@5: 35.34 -> 19.56 | Hurt |
| add CNN score prediction after SoftArgMax (ablation) | SuperPoint + MNN · MegaDepth relative pose | AUC@5: 19.56 -> 36.71 | Improved |
| add descriptor guidance (without mean feature) (ablation) | SuperPoint + MNN · MegaDepth relative pose | AUC@5: 36.71 -> 36.57 | No clear effect |
| add mean feature (average of both views' descriptors) (ablation) | SuperPoint + MNN · MegaDepth relative pose | AUC@5: 36.57 -> 37.16 | Improved |
| apply the refinement on top of SIFT | SIFT + MNN · MegaDepth relative pose | AUC@5: 9.53 -> 9.49 | Hurt |
| apply the refinement on top of DeDoDe | DeDoDe + DSM · MegaDepth relative pose | AUC@5: 48.83 -> 52.57 | Improved |
| apply the refinement on SuperPoint + MNN for keypoint localization | SuperPoint + MNN · HPatches | MMA@1px: 27% -> 33% | Improved |
Closest papers by meaning
- P2-Net: Joint Description and Detection of Local Features for Pixel and Point Matching (ICCV 2021 · similarity 0.87)
- ASLFeat: Learning Local Features of Accurate Shape and Localization (CVPR 2020 · similarity 0.87)
- R2D2: Reliable and Repeatable Detector and Descriptor (NeurIPS 2019 · similarity 0.86)
- Reinforced Feature Points: Optimizing Feature Detection and Description for a High-Level Task (CVPR 2020 · similarity 0.85)
- PolarNet: Learning to Optimize Polar Keypoints for Keypoint Based Object Detection (ICLR 2021 · similarity 0.85)
- Learning to Produce Semi-dense Correspondences for Visual Localization (CVPR 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.
