Research map / Detection and segmentation / CNN features
Towards Balanced Learning for Instance Recognition
Jiangmiao Pang, Kai Chen, Qi Li, Zhihai Xu, Huajun Feng, Jianping Shi, Wanli Ouyang, Dahua Lin
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 IoU-balanced sampling (ablation) | Faster R-CNN R50-FPN · COCO val-2017 | AP: 35.9 -> 36.8 | Improved |
| add balanced feature pyramid (ablation) | Faster R-CNN with IoU-balanced sampling · COCO val-2017 | AP: 36.8 -> 37.7 | Improved |
| add LGR-L1 loss (ablation) | Libra R-CNN · COCO val-2017 | AP: 37.7 -> 38.5 | Improved |
| add LGR-CE loss (ablation) | Libra R-CNN · COCO val-2017 | AP: 38.5 -> 39.0 | Improved |
| vary number of IoU bins K (2, 3, 5) (ablation) | IoU-balanced sampling · COCO val-2017 | AP: 36.7 / 36.8 / 36.7 | No clear effect |
| balance positive samples per ground truth (ablation) | Faster R-CNN · COCO val-2017 | AP: 35.9 -> 36.1 | No clear effect |
| add non-local refinement to balanced semantic features (ablation) | Balanced feature pyramid · COCO val-2017 | AP: 36.3 -> 36.8 | Improved |
| add balanced semantic features to PAFPN (ablation) | PAFPN · COCO val-2017 | AP: 36.3 -> 37.2 | Improved |
| raise localization loss weight from 1.5 to 2.0 (ablation) | Faster R-CNN · COCO val-2017 | AP: 36.4 -> 36.2 | Hurt |
| replace balanced L1 (alpha 0.5, gamma 1.0) with gamma 1.5 (ablation) | Balanced L1 loss · COCO val-2017 | AP: 36.5 -> 37.2 | Improved |
| use plain L1 loss instead of balanced L1 (ablation) | Faster R-CNN · COCO val-2017 | AP: 36.6 vs 37.2 | Hurt |
| add bucketed class re-weighting (BCR) (ablation) | Libra R-CNN · LVIS val | APr: 14.5 -> 16.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Cascade R-CNN — Cascade R-CNN: Delving Into High Quality Object Detection (CVPR 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Libra R-CNN: Towards Balanced Learning for Object Detection (CVPR 2019 · similarity 0.90)
- A Re-Balancing Strategy for Class-Imbalanced Classification Based on Instance Difficulty (CVPR 2022 · similarity 0.86)
- BaCon: Boosting Imbalanced Semi-supervised Learning via Balanced Feature-Level Contrastive Learning (AAAI 2024 · similarity 0.82)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.82)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.81)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.81)
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.
