Research map / Label-efficient and robust learning / Recognition and long-tail
S2DNet: Learning Image Features for Accurate Sparse-to-Dense Matching
Hugo Germain, Guillaume Bourmaud, Vincent Lepetit
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 sparse-to-dense instead of sparse-to-sparse matching (ablation) | S2DNet + SuperPoint · HPatches | MMA@1px: 0.373 -> 0.563 | Improved |
| use sparse-to-dense instead of sparse-to-sparse matching (ablation) | S2DNet + D2-Net detector · HPatches | MMA@1px: 0.118 -> 0.467 | Improved |
| add confidence thresholding tau = 0.20 (ablation) | S2DNet + SuperPoint · HPatches | MMA@1px: 0.469 -> 0.563 | Improved |
| add confidence thresholding tau = 0.20 (ablation) | S2DNet + Harris · HPatches | MMA@1px: 0.441 -> 0.511 | Improved |
| use SuperPoint instead of Harris detector for source keypoints (ablation) | S2DNet · HPatches | MMA@1px: 0.511 -> 0.563 | Improved |
| use sparse-to-dense with D2-Net and R2D2 features under illumination change (ablation) | D2-Net and R2D2 · HPatches illumination | MMA | Improved |
| use sparse-to-dense with D2-Net and R2D2 features under viewpoint change (ablation) | D2-Net and R2D2 · HPatches viewpoint | MMA | Hurt |
| run weakly supervised S2DHM in sparse-to-sparse setting (ablation) | S2DHM · HPatches | MMA | Improved |
| replace weakly supervised features with supervised classification-trained S2DNet features | S2DHM · Aachen Night | recall 0.25m/2deg: 30.6 -> 46.9 | Improved |
| use S2DNet instead of D2-Net on day localization | D2-Net · RobotCar Seasons Day-All | recall 0.25m/2deg: 54.5 -> 53.9 | No clear effect |
| use S2DNet instead of D2-Net on nighttime RobotCar finest threshold | D2-Net · RobotCar Night-All | recall 0.25m/2deg: 20.4 -> 14.5 | Hurt |
| use S2DNet instead of D2-Net on InLoc direct PE medium and coarse thresholds | D2-Net · InLoc Direct PE | recall 5m/10deg: 48.6 -> 48.5 | No clear effect |
Closest papers by meaning
- NCNet: Neighbourhood Consensus Networks for Estimating Image Correspondences (TPAMI 2020 · similarity 0.87)
- NM-Net: Mining Reliable Neighbors for Robust Feature Correspondences (CVPR 2019 · similarity 0.87)
- Neighbourhood Consensus Networks (NeurIPS 2018 · similarity 0.87)
- Correspondence Transformers With Asymmetric Feature Learning and Matching Flow Super-Resolution (CVPR 2023 · similarity 0.87)
- Reinforced Feature Points: Optimizing Feature Detection and Description for a High-Level Task (CVPR 2020 · similarity 0.86)
- Deep Matching Prior: Test-Time Optimization for Dense Correspondence (ICCV 2021 · 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.
