Research map / Detection and segmentation / Weakly supervised detection and segmentation
BoxSnake: Polygonal Instance Segmentation with Box Supervision
Rui Yang, Lin Song, Yixiao Ge, Xiu Li
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 distance-aware pairwise losses on top of the point-based unary loss (ablation) | BoxSnake · COCO val2017 | mask AP: 23.9 -> 31.1 | Improved |
| add local pairwise term (unary + local) (ablation) | BoxSnake · COCO val2017 | mask AP: 23.9 -> 30.8 | Improved |
| add global pairwise term (unary + global) (ablation) | BoxSnake · COCO val2017 | mask AP: 23.9 -> 29.8 | Improved |
| use CIoU/GIoU point-based unary loss instead of Dice on P3 (ablation) | BoxSnake · COCO val2017 | mask AP: 19.3 -> 23.9 | Improved |
| add clipping strategy instead of scaling polygon to P2 (ablation) | BoxSnake · COCO val2017 | mask AP: 29.8 -> 31.1 | Improved |
| increase pairwise-loss resolution from P3 to P2 size (ablation) | BoxSnake · COCO val2017 | mask AP: 29.6 -> 29.8 | No clear effect |
| vary local pairwise window size and dilation (ablation) | BoxSnake · COCO val2017 | mask AP: 30.8 / 30.9 / 31.1 | No clear effect |
| use ellipse instead of square initial polygon (ablation) | BoxSnake · COCO val2017 | mask AP: 30.7 -> 31.1 | Improved |
| use proposed local pairwise loss instead of CRF loss (ablation) | BoxSnake · COCO val2017 | mask AP: 30.2 -> 31.1 | Improved |
| set sigmoid temperature tau to 0.1 rather than 0.5 (ablation) | BoxSnake · COCO val2017 | mask AP: 29.9 -> 31.1 | Improved |
| polygon output with box supervision instead of mask output (BoxInst) | BoxSnake · Cityscapes val | mask AP: 22.4 -> 26.3 | Improved |
| use BoxSnake polygon head instead of BoxInst mask head | box-supervised instance segmentation · COCO val2017 (R50, 1x) | mask AP: 30.7 -> 31.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- BoxInst — BoxInst: High-Performance Instance Segmentation With Box Annotations (CVPR 2021) · better on 3 of 4 shared results
- DiscoBox — DiscoBox: Weakly Supervised Instance Segmentation and Semantic Correspondence From Box Supervision (ICCV 2021) · 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
- PolyTransform: Deep Polygon Transformer for Instance Segmentation (CVPR 2020 · similarity 0.90)
- BoxInst: High-Performance Instance Segmentation With Box Annotations (CVPR 2021 · similarity 0.90)
- Sketchy Bounding-box Supervision for 3D Instance Segmentation (CVPR 2025 · similarity 0.89)
- Extreme Point Supervised Instance Segmentation (CVPR 2024 · similarity 0.87)
- BSNet: Box-Supervised Simulation-assisted Mean Teacher for 3D Instance Segmentation (CVPR 2024 · similarity 0.86)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · 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.
