Research map / Label-efficient and robust learning / Recognition and long-tail
Relieving Long-Tailed Instance Segmentation via Pairwise Class Balance
Yin-Yin He, Peizhen Zhang, Xiu-Shen Wei, Xiangyu Zhang, Jian Sun
Check 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 PCB to Seesaw | Seesaw on LVIS v0.5 · LVIS v0.5 val, R-50-FPN | AP_r: 19.9 -> 23.4 | Improved |
| Add PCB to Seesaw | Seesaw on LVIS v1.0 · LVIS v1.0 val, R-50-FPN | AP: 26.8 -> 27.2 | Improved |
| Add PCB regularization (ablation) | RFS on LVIS v0.5 · LVIS v0.5 val, R-50-FPN | AP_r: 16.0 -> 18.9 | Improved |
| Replace PCB regularization with online label smoothing (ablation) | RFS on LVIS v0.5 · LVIS v0.5 val, R-50-FPN | AP_r: 16.0 -> 15.4 | Hurt |
| Add iterative learning paradigm without PCB (ablation) | RFS on LVIS v0.5 · LVIS v0.5 val, R-50-FPN | AP: 25.6 -> 26.5 | Improved |
| Replace the iterative paradigm with deep supervision (ablation) | PCB regularization with RFS · LVIS v0.5 val, R-50-FPN | AP: 27.7 -> 26.9 | Hurt |
| Increase EMA momentum (ablation) | PCB with RFS · LVIS v0.5 val, R-50-FPN | AP_r: 19.3 -> 22.1 | Improved |
| Increase PCB regularization coefficient from 0.4 to 1.0 (ablation) | PCB with RFS · LVIS v0.5 val, R-50-FPN | AP: 27.7 -> 26.6 | Hurt |
| Use prediction-dependent iterative refinement instead of the CrowdDet refinement module (ablation) | PCB regularization with RFS · LVIS v0.5 val, R-50-FPN | AP: 26.4 -> 27.7 | Improved |
| Add PCB to CE | CE on ImageNet-LT · ImageNet-LT, ResNeXt-50 | Few accuracy: 7.44 -> 21.97 | Improved |
Closest papers by meaning
- The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation (ECCV 2020 · similarity 0.90)
- Seesaw Loss for Long-Tailed Instance Segmentation (CVPR 2021 · similarity 0.88)
- DropLoss for Long-Tail Instance Segmentation (AAAI 2021 · similarity 0.88)
- Generalized Class Discovery in Instance Segmentation (AAAI 2025 · similarity 0.87)
- Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.87)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · 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.
