Research map / Label-efficient and robust learning / Recognition and long-tail
FASA: Feature Augmentation and Sampling Adaptation for Long-Tailed Instance Segmentation
Yuhang Zang, Chen Huang, Chen Change Loy
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 |
|---|---|---|---|
| add adaptive feature augmentation (FA) (ablation) | Mask R-CNN RFS · LVIS v1.0 val | AP: 20.8 -> 22.3 | Improved |
| add adaptive feature sampling (FS) on top of FA (ablation) | Mask R-CNN RFS + FA · LVIS v1.0 val | AP: 22.3 -> 23.7 | Improved |
| add adaptive FS on top of FA, rare classes (ablation) | Mask R-CNN RFS + FA · LVIS v1.0 val | AP rare: 12.7 -> 17.8 | Improved |
| use adaptive sampling instead of static s=5 (ablation) | FASA · LVIS v1.0 val | AP: 22.3 -> 23.7 | Improved |
| static scale s=5 vs s=1 on frequent classes (ablation) | FASA static sampling · LVIS v1.0 val | AP_r and AP_f: s=5 best for AP_r, s=1 best for AP_f | Mixed |
| add FASA to vanilla Mask R-CNN | Mask R-CNN · LVIS v1.0 val | AP: 19.3 -> 22.6 | Improved |
| add FASA to RFS | RFS · LVIS v1.0 val | AP: 22.8 -> 24.1 | Improved |
| add FASA to EQL | EQL · LVIS v1.0 val | AP: 22.1 -> 24.4 | Improved |
| add FASA to Seesaw loss | Seesaw + RFS · LVIS v1.0 val | AP: 26.4 -> 27.5 | Improved |
| add FASA to Mask R-CNN with ResNet-101 | Mask R-CNN RFS R101 · LVIS v1.0 val | AP: 24.4 -> 26.3 | Improved |
| add FASA to Cascade Mask R-CNN with Seesaw | Cascade Mask R-CNN Seesaw R101 · LVIS v1.0 val | AP: 30.1 -> 31.5 | Improved |
| add FASA to Mask R-CNN on COCO-LT | Mask R-CNN · COCO-LT minival | AP mask: 18.7 -> 23.4 | Improved |
Closest papers by meaning
- SAFA: Sample-Adaptive Feature Augmentation for Long-Tailed Image Classification (ECCV 2022 · similarity 0.89)
- Learning with Free Object Segments for Long-Tailed Instance Segmentation (ECCV 2022 · similarity 0.88)
- The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation (ECCV 2020 · similarity 0.88)
- Feature Space Augmentation for Long-Tailed Data (ECCV 2020 · similarity 0.88)
- Learning Only When It Matters: Cost-Aware Long-Tailed Classification (AAAI 2024 · similarity 0.85)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.84)
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.
