Research map / Detection and segmentation / CNN features
Multi-Scale and Cross-Scale Contrastive Learning for Semantic Segmentation
Theodoros Pissas, Claudio S. Ravasio, Lyndon Da Cruz, Christos Bergeles
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 single-scale contrastive loss (ablation) | HRNet HR48v2 with CE · Cityscapes val | mIoU: 79.1 -> 80.0 | Improved |
| extend contrastive loss to multiple scales (ablation) | HRNet HR48v2 · Cityscapes val | mIoU: 80.0 -> 80.7 | Improved |
| add cross-scale contrastive loss (1 scale pair) to multi-scale loss (ablation) | HRNet HR48v2 · Cityscapes val | mIoU: 80.7 -> 81.2 | Improved |
| use 2 cross-scale pairs instead of 1 (ablation) | HRNet HR48v2 · Cityscapes val | mIoU: 81.2 -> 81.4 | Improved |
| add cross-scale loss to multi-scale loss (ablation) | UPerNet Swin-T · ADE20K val | mIoU: 45.1 -> 45.8 | Improved |
| add multi-scale contrastive loss (ablation) | UPerNet Swin-T · ADE20K val | mIoU: 44.5 -> 45.1 | Improved |
| add balanced anchor sampling instead of using all anchors, single-scale loss (ablation) | HRNet · Cityscapes val | mIoU: 79.4 -> 80.2 | Improved |
| use anchor sampling for multi-scale and cross-scale losses (ablation) | HRNet · Cityscapes val | memory: OOM -> 9.7 GB/GPU | Improved |
| use decreasing per-scale weights 1.0 0.7 0.4 0.1 instead of equal weights (ablation) | HRNet · Cityscapes val | mIoU: 81.8 -> 82.2 | Improved |
| apply loss on backbone instead of FPN outputs (ablation) | UPerNet Swin-S · ADE20K val | mIoU: 49.0 -> 47.9 | Hurt |
| add contrastive loss to CE | OCRNet ResNet-50 · CaDIS task 3 rare classes | IoU: 67.58 -> 73.09 | Improved |
| add contrastive loss to CE | UPerNet Swin-L · ADE20K val multi-scale | mIoU: 53.5 -> 53.3 | Mixed |
Closest papers by meaning
- Contrastive Learning for Label Efficient Semantic Segmentation (ICCV 2021 · similarity 0.92)
- Looking Beyond Single Images for Contrastive Semantic Segmentation Learning (NeurIPS 2021 · similarity 0.91)
- Region-Aware Contrastive Learning for Semantic Segmentation (ICCV 2021 · similarity 0.91)
- Exploring Cross-Image Pixel Contrast for Semantic Segmentation (ICCV 2021 · similarity 0.90)
- Contextrast: Contextual Contrastive Learning for Semantic Segmentation (CVPR 2024 · similarity 0.89)
- Multi-Scale Representations by Varying Window Attention for Semantic Segmentation (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.
