Research map / Detection and segmentation / Open-set and few-shot detection
The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation
Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, Jiashi Feng
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 classification-head calibration with bi-level sampling (ablation) | Mask R-CNN r50-ag · LVIS v0.5 val | mask AP1: 0.0 -> 13.2 | Improved |
| add classification-head calibration with bi-level sampling (ablation) | Mask R-CNN r50 (class-wise) · LVIS v0.5 val | mask AP4: 29.6 -> 25.3 | Hurt |
| add dual head inference on top of calibration (ablation) | Mask R-CNN r50 · LVIS v0.5 val | mask AP: 22.3 -> 23.4 | Improved |
| add dual head inference on top of calibration (ablation) | Mask R-CNN r50-ag · LVIS v0.5 val | mask AP: 21.2 -> 22.2 | Improved |
| combine heads by selecting by class instance count instead of averaging predictions (ablation) | SimCal dual head · LVIS v0.5 val | mask AP: 20.3 -> 21.1 | Improved |
| initialize calibration head from the original 3-layer head instead of random initialization (ablation) | SimCal head · LVIS v0.5 val (HTC) | mask AP1 / AP2: +4.1 AP1, +4.3 AP2 | Improved |
| vary head/tail boundary T from 10 to 1000 (ablation) | SimCal dual head · LVIS v0.5 val | AP: T 90 to 500 near optimal (T = 300) | No clear effect |
| use class-agnostic heads instead of class-wise heads with calibration and dual head (ablation) | SimCal · LVIS v0.5 val | mask AP1: 10.2 -> 13.3 | Improved |
| add loss re-weighting during training | Mask R-CNN r50 · LVIS v0.5 val | mask AP4: 29.6 -> 25.1 | Hurt |
| add image-level repeat sampling during training | Mask R-CNN r50 · LVIS v0.5 val | mask AP: 20.7 -> 23.2 | Improved |
| add decoupled calibration and dual head to existing loss methods CM, LR, FL (ablation) | Mask R-CNN r50 · LVIS v0.5 val | mask AP1: 2.6 -> 4.6; 6.0 -> 6.9; 6.2 -> 7.1 | Improved |
| add decoupled calibration to image-level repeat sampling (ablation) | Mask R-CNN r50 · LVIS v0.5 val | mask AP: 23.2 -> 22.7 | Hurt |
Closest papers by meaning
- Unsupervised Discovery of the Long-Tail in Instance Segmentation Using Hierarchical Self-Supervision (CVPR 2021 · similarity 0.92)
- On Model Calibration for Long-Tailed Object Detection and Instance Segmentation (NeurIPS 2021 · similarity 0.91)
- Learning with Free Object Segments for Long-Tailed Instance Segmentation (ECCV 2022 · similarity 0.91)
- Relieving Long-Tailed Instance Segmentation via Pairwise Class Balance (CVPR 2022 · similarity 0.90)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.89)
- The Devil is in the Object Boundary: Towards Annotation-free Instance Segmentation using Foundation Models (ICLR 2024 · similarity 0.87)
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.
