Research map / Detection and segmentation / CNN features
Rethinking Semantic Segmentation: A Prototype View
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 |
|---|---|---|---|
| Replace parametric class prediction with non-learnable prototypes | SegFormer · ADE20K val | mIoU: 50.9 -> 51.7 | Helped |
| Replace parametric class prediction with non-learnable prototypes | SegFormer · Cityscapes val | mIoU: 80.7 -> 81.3 | Helped |
| Replace parametric class prediction with non-learnable prototypes | SegFormer · COCO-Stuff test | mIoU: 42.5 -> 43.3 | Helped |
| Add pixel-prototype contrastive loss to cross-entropy (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 45.0 -> 45.9 | Helped |
| Add pixel-prototype distance loss to cross-entropy (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 45.0 -> 45.4 | Helped |
| Use all three losses together (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 45.0 -> 46.4 | Helped |
| Increase prototypes per class from 1 to 10 (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 45.5 -> 46.4 | Helped |
| Increase prototypes per class beyond 10 (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 46.4 -> 46.5 -> 46.4 | No clear effect |
| Use slower prototype updates (μ=0.999 rather than μ=0) (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 44.9 -> 46.4 | Helped |
| Use cosine distance instead of standard Euclidean distance (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 45.7 -> 46.4 | Helped |
| Use Huberized distance instead of standard Euclidean distance (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 45.7 -> 45.2 | No clear effect |
| Vary λ1 across the tested values (ablation) | Prototype segmentation, MiT-B2 · ADE20K val | mIoU: 46.1, 46.3, 46.4, 46.4, 46.2, 46.3 | No clear effect |
Closest papers by meaning
- CLIP-S4: Language-Guided Self-Supervised Semantic Segmentation (CVPR 2023 · similarity 0.89)
- Hierarchical Prototype Learning for Semantic Segmentation (ICLR 2026 · similarity 0.89)
- Pixelwise Instance Segmentation With a Dynamically Instantiated Network (CVPR 2017 · similarity 0.89)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.88)
- Generalized Few-Shot Semantic Segmentation (CVPR 2022 · similarity 0.88)
- Bayesian Decomposition and Semantic Completion for Few-shot Semantic Segmentation (CVPR 2026 · similarity 0.88)
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.