Research map / Detection and segmentation / CNN features
DR Loss: Improving Object Detection by Distributional Ranking
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 focal loss with DR loss | RetinaNet with ResNet-101-FPN · COCO test-dev | AP: 39.1 -> 41.7 | Helped |
| Replace focal loss with DR loss | RetinaNet · PASCAL VOC2007 test | AP: 39.5 -> 41.2 | Helped |
| Use a smooth DR ranking loss instead of hinge loss (ablation) | RetinaNet on COCO minival · COCO minival | AP: 35.8 -> 37.4 | Helped |
| Use logistic loss with L=6 instead of focal loss (ablation) | RetinaNet, ResNet-50 · COCO minival | AP: 36.1 -> 37.4 | Helped |
| Set logistic smoothness parameter L=10 instead of L=6 (ablation) | DR Loss on RetinaNet · COCO minival | AP: 37.4 -> 36.8 | Hurt |
| Set positive-distribution parameter h+ to 20 instead of e (ablation) | DR Loss on RetinaNet · COCO minival | AP: 37.4 -> 35.6 | Hurt |
| Set negative-distribution parameter h- to 20 instead of 3.5 (ablation) | DR Loss on RetinaNet · COCO minival | AP: 37.4 -> 36.6 | Hurt |
| Rank all positive-negative pairs instead of distribution expectations (ablation) | Ranking loss on RetinaNet · COCO minival | AP: 37.4 -> 12.9 | Hurt |
| Use negative-only distributional ranking instead of ranking both distributions (ablation) | DR Loss on RetinaNet · COCO minival | AP: 37.0 -> 37.4 | Hurt |
| Increase the confidence threshold from 0.05 to 0.5 (ablation) | Focal loss on RetinaNet · COCO minival | AP: 36.1 -> 28.4 | Hurt |
| Increase the confidence threshold from 0.05 to 0.5 (ablation) | DR Loss on RetinaNet · COCO minival | AP: 37.4 -> 37.2 | No clear effect |
| Use DR loss instead of focal loss | RetinaNet, ResNet-50, input scale 800 · COCO test-dev | AP: 35.7 -> 37.6 | Helped |
Closest papers by meaning
- RankDetNet: Delving Into Ranking Constraints for Object Detection (CVPR 2021 · similarity 0.91)
- Revisiting AP Loss for Dense Object Detection: Adaptive Ranking Pair Selection (CVPR 2022 · similarity 0.89)
- RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses (CVPR 2024 · similarity 0.86)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.86)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.86)
- Supervision Interpolation via LossMix: Generalizing Mixup for Object Detection and Beyond (AAAI 2024 · similarity 0.85)
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.