Research map / Detection and segmentation / CNN features
Instance Segmentation With Mask-Supervised Polygonal Boundary Transformers
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| regress polygons with mask-space supervision and differentiable rasterization | BoundaryFormer with R50-FPN R-CNN · MS-COCO val | AP: 35.2 -> 36.1 | Helped |
| use BoundaryFormer instead of Mask R-CNN for a COCO-initialized model | R50-FPN R-CNN · Cityscapes val | AP: 36.5 -> 38.3 | Helped |
| replace annotator polygons with polygons generated from masks (ablation) | DANCE · MS-COCO val | AP: 34.5 -> 23.1 | Hurt |
| remove the half-pixel coordinate adjustment for rasterizer alignment (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 35.3 | Hurt |
| remove global pooling-like operation in the RoI-less setting (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 34.2 | Hurt |
| increase rasterization smoothness to τ = 1.0 (ablation) | BoundaryFormer · MS-COCO val | AP: 35.6 | Hurt |
| reduce rasterization resolution to 14 × 14 (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 34.5 | Hurt |
| use rasterization resolution 20 × 20 instead of 64 × 64 (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 35.5 | Hurt |
| use rasterization resolution 40 × 40 instead of 64 × 64 (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 35.9 | Hurt |
| increase rasterization resolution from 64 × 64 to 80 × 80 (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 36.1 | No clear effect |
| use 2 refinement layers with 32 initial and 64 final points instead of 4 layers with 8/64 points (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 35.0 | Hurt |
| use 3 refinement layers with 16 initial and 64 final points instead of 4 layers with 8/64 points (ablation) | BoundaryFormer · MS-COCO val | AP: 36.1 -> 35.7 | Hurt |
Closest papers by meaning
- Look Closer To Segment Better: Boundary Patch Refinement for Instance Segmentation (CVPR 2021 · similarity 0.91)
- PolyTransform: Deep Polygon Transformer for Instance Segmentation (CVPR 2020 · similarity 0.91)
- Boundary-Aware Instance Segmentation (CVPR 2017 · similarity 0.90)
- EFDTR: Learnable Elliptical Fourier Descriptor Transformer for Instance Segmentation (ICML 2025 · similarity 0.88)
- PEM: Prototype-based Efficient MaskFormer for Image Segmentation (CVPR 2024 · similarity 0.87)
- PolarNeXt: Rethink Instance Segmentation with Polar Representation (CVPR 2025 · 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.