Research map / Learning theory and algorithms / Prediction and decision losses
Optimization of Rank Losses for Image Retrieval
Elias Ramzi, Nicolas Audebert, Clément Rambour, André de Melo Araújo, Xavier Bitot, Nicolas Thome
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 |
|---|---|---|---|
| replace sigmoid rank with SupRank (ablation) | Smooth-AP · SOP | mAP@R: 54.3 -> 54.8 | Improved |
| replace sigmoid rank with SupRank (ablation) | Smooth-AP · iNaturalist | mAP@R: 26.5 -> 27.5 | Improved |
| add proxy-based decomposability loss L_DG* (ablation) | Sup-AP · iNaturalist | R@1: 68.9 -> 71.8 | Improved |
| add proxy-based decomposability loss L_DG* (ablation) | Sup-R@k · iNaturalist | R@1: 68.2 -> 71.9 | Improved |
| use proxy-based L_DG* instead of pair-based L_DG (ablation) | ROADMAP · iNaturalist | R@1: 69.1 -> 71.8 | Improved |
| use proxy-based L_DG* instead of pair-based L_DG (ablation) | ROADMAP · SOP | R@1: 81.7 -> 81.9 | No clear effect |
| reduce batch size from 384 to 32 when adding L_DG* (ablation) | Sup-AP · iNaturalist | relative mAP@R increase: 3.5 -> 22.8 | Improved |
| increase lambda from 0.1 to 0.3 (ablation) | ROADMAP · iNaturalist | R@1: 71.8 -> 72.1 | Improved |
| increase slope rho in SupRank from 0 to above 1 (ablation) | Sup-AP · iNaturalist | mAP@R | Improved |
| use ROADMAP instead of cross-batch memory XBM | XBM · iNaturalist | R@1: 59.3 -> 71.8 | Improved |
| optimize H-AP (HAPPIER) instead of CSL | CSL · SOP | H-AP: 52.8 -> 59.4 | Improved |
| optimize H-AP with hierarchy instead of fine-grained Smooth-AP | Smooth-AP · SOP | H-AP: 42.9 -> 59.4 | Improved |
Closest papers by meaning
- Robust and Decomposable Average Precision for Image Retrieval (NeurIPS 2021 · similarity 0.92)
- Hierarchical Average Precision Training for Pertinent Image Retrieval (ECCV 2022 · similarity 0.91)
- Learning With Average Precision: Training Image Retrieval With a Listwise Loss (ICCV 2019 · similarity 0.90)
- Efficient Optimization for Rank-Based Loss Functions (CVPR 2018 · similarity 0.90)
- RankMI: A Mutual Information Maximizing Ranking Loss (CVPR 2020 · similarity 0.90)
- Threshold-Consistent Margin Loss for Open-World Deep Metric Learning (ICLR 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.
