Research map / Detection and segmentation / Weakly supervised detection and segmentation
RIPE: Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction
Johannes Künzel, Anna Hilsmann, Peter Eisert
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 Tokyo 24/7 day-night pairs (80/20 mix) to training (ablation) | RIPE · Aachen Day-Night v1.1 (Night, 5m/10deg) | Localization accuracy (%): 79.1 -> 83.8 | Improved |
| add Tokyo 24/7 day-night pairs (80/20 mix) to training (ablation) | RIPE · Aachen Day-Night v1.1 (Night, .25m/2deg) | Localization accuracy (%): 52.9 -> 60.2 | Improved |
| add Tokyo 24/7 data to training (ablation) | RIPE · Aachen Day-Night v1.1 (Day, 5m/10deg) | Localization accuracy (%): 93.1 -> 92.4 | Hurt |
| increase Tokyo share to 60% of training data (ablation) | RIPE · Aachen Day-Night v1.1 (Day, .25m/2deg) | Localization accuracy (%): 81.6 -> 71.8 | Hurt |
| add descriptor loss (psi>0) (ablation) | RIPE · IMC2020 subset | AUC@5deg: failed to train at psi=0 -> 63.48 at psi=5 | Improved |
| raise descriptor loss weight from 0.05 to 5 (ablation) | RIPE · IMC2020 subset | AUC@5deg: 61.94 -> 63.48 | Improved |
| raise descriptor loss weight from 5 to 50 (ablation) | RIPE · IMC2020 subset | AUC@5deg: 63.48 -> 60.65 | Hurt |
| add keypoint-count regularization eps=-7e-7 (ablation) | RIPE · IMC2020 subset | AUC@5deg: 61.0 -> 63.48 | Improved |
| use too large regularization (eps=-7e-6 or stronger) (ablation) | RIPE · IMC2020 subset | AUC@5deg: training failed | Hurt |
| replace hyper-column descriptors with bilinear upsampling of final encoder layer (ablation) | RIPE · IMC2020 subset | training: training failed | Hurt |
| use RIPE (unlabeled pairs) instead of ALIKED on MegaDepth-1500 | RIPE · MegaDepth-1500 | AUC@5deg: 55.11 vs 56.71 | Mixed |
| use RIPE instead of DeDoDe-B | RIPE · MegaDepth-1500 | AUC@5deg: 55.02 -> 55.11 | Improved |
Closest papers by meaning
- Reinforced Feature Points: Optimizing Feature Detection and Description for a High-Level Task (CVPR 2020 · similarity 0.88)
- From Pairs to Sequences: Track-Aware Policy Gradients for Keypoint Detection (CVPR 2026 · similarity 0.87)
- S-TREK: Sequential Translation and Rotation Equivariant Keypoints for Local Feature Extraction (ICCV 2023 · similarity 0.86)
- ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation (CVPR 2024 · similarity 0.86)
- DISK: Learning local features with policy gradient (NeurIPS 2020 · similarity 0.85)
- Incremental Object Keypoint Learning (CVPR 2025 · similarity 0.84)
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.
