Research map / Detection and segmentation / CNN features
Optimizing Rank-Based Metrics With Blackbox Differentiation
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 cross-entropy with RaMBO AP losses | Faster R-CNN · Pascal VOC 07 test, ResNet50 | AP50: 74.2 -> 75.7 | Helped |
| Replace cross-entropy with RaMBO AP losses | Faster R-CNN · Pascal VOC 07 test, ResNet50, training on 07+12 | AP50: 80.4 -> 81.4 | Helped |
| Replace cross-entropy with RaMBO AP losses | Faster R-CNN · Pascal VOC 07 test, ResNet101, training on 07+12 | AP50: 82.4 -> 82.9 | Helped |
| Replace cross-entropy with RaMBO AP losses | Faster R-CNN · Pascal VOC 07 test, ResNeXt101 32x4d, training on 07+12 | AP50: 83.2 -> 83.6 | Helped |
| Use RaMBO loglog rather than log rank weighting | RaMBO retrieval loss · Stanford Online Products | R@1: 77.8 -> 78.6 | Helped |
| Use RaMBO loglog rather than log rank weighting | RaMBO retrieval loss · CUB-200-2011 | R@1: 63.5 -> 64.0 | No clear effect |
| Use RaMBO loglog rather than log rank weighting | RaMBO retrieval loss · In-shop Clothes | R@1: 88.1 -> 86.3 | Hurt |
| Add batch memory (ablation) | RaMBO retrieval loss · CUB-200-2011 | R@1: 62.5 -> 64.0 | Helped |
| Add batch memory (ablation) | RaMBO retrieval loss · In-shop Clothes | R@1: 87.0 -> 88.1 | Helped |
| Add batch memory (ablation) | RaMBO retrieval loss · Stanford Online Products | R@1: 72.4 -> 78.6 | Helped |
| Add score margin (ablation) | RaMBO retrieval loss · CUB-200-2011 | R@1: 63.2 -> 64.0 | Helped |
| Add score margin (ablation) | RaMBO AP loss · Pascal VOC 07 test, Faster R-CNN ResNet50 | AP50: 74.6 -> 75.7 | Helped |
Closest papers by meaning
- MetricOpt: Learning To Optimize Black-Box Evaluation Metrics (CVPR 2021 · similarity 0.89)
- Deep Metric Learning to Rank (CVPR 2019 · similarity 0.88)
- RankDetNet: Delving Into Ranking Constraints for Object Detection (CVPR 2021 · similarity 0.88)
- RankMI: A Mutual Information Maximizing Ranking Loss (CVPR 2020 · similarity 0.86)
- Revisiting AP Loss for Dense Object Detection: Adaptive Ranking Pair Selection (CVPR 2022 · similarity 0.85)
- A General Online Algorithm for Optimizing Complex Performance Metrics (ICML 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.