Research map / Detection and segmentation / CNN features
Learning to Downsample for Segmentation of Ultra-High Resolution Images
Chen Jin, Ryutaro Tanno, Thomy Mertzanidou, Eleftheria Panagiotaki, Daniel C. Alexander
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 |
|---|---|---|---|
| Learn sampling locations jointly with segmentation and add edge-loss regularization | Deformation module · Cityscapes, DeepGlobe, and PCa-Histo | mIoU: 3% to 10% higher absolute mIoU over the "uniform" baseline | Improved |
| Add edge-loss regularization to single segmentation loss (ablation) | Deformation module · Cityscapes, DeepGlobe, and PCa-Histo | mIoU | Improved |
| Use learned non-uniform downsampling instead of uniform downsampling | Segmentation model · Cityscapes fair comparison | mIoU: upto 4% in absolute mIoU | Improved |
| Use learned non-uniform downsampling instead of uniform downsampling | Segmentation model · Cityscapes, DeepGlobe, and PCa-Histo | computational cost at comparable performance: 33%, 90% and 70% | Improved |
| Use single segmentation loss without edge-loss regularization (ablation) | Deformation module · Cityscapes, DeepGlobe, and PCa-Histo | mIoU versus uniform downsampling: 2/3 tested datasets | Mixed |
| Set Gaussian kernel standard deviation from 15 to 26 or 32 (ablation) | Deformation module · Cityscapes (1024² to 96² pixels) | mIoU: 0.35 -> 0.34; 0.35 -> 0.32 | Hurt |
| Set Gaussian kernel standard deviation from 20 to 25 (ablation) | Deformation module · DeepGlobe (2448² to 100² pixels) | mIoU: 0.43 -> 0.45 | Improved |
| Set deformation-module kernel size to 5×5 or 7×7 instead of 3×3 (ablation) | Deformation module · Cityscapes (1024² to 96² pixels) | mIoU: 0.36 -> 0.35; 0.36 -> 0.35 | Hurt |
| Set deformation-module kernel size to 7×7 instead of 3×3 (ablation) | Deformation module · DeepGlobe (2448² to 100² pixels) | mIoU: 0.45 -> 0.46 | Improved |
| Vary deformation-module low-resolution input size (ablation) | Deformation module · Cityscapes (1024² to 64² pixels) | mIoU: 0.35, 0.35, 0.34, 0.33, 0.36 | Mixed |
| Vary deformation-module low-resolution input size (ablation) | Deformation module · DeepGlobe (2448² to 50² pixels) | mIoU: 0.44, 0.42, 0.45, 0.42, 0.43 | Mixed |
| Increase edge-loss weight from γ=50 to γ=100 (ablation) | Deformation module · Cityscapes (1024² to 64²) | mIoU: 0.33 -> 0.35 | Improved |
Closest papers by meaning
- Efficient Segmentation: Learning Downsampling Near Semantic Boundaries (ICCV 2019 · similarity 0.88)
- Upsample Anything: A Simple and Hard to Beat Baseline for Feature Upsampling (CVPR 2026 · similarity 0.87)
- Boosting the Dual-Stream Architecture in Ultra-High Resolution Segmentation with Resolution-Biased Uncertainty Estimation (CVPR 2025 · similarity 0.84)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.84)
- Native Segmentation Vision Transformers (NeurIPS 2025 · similarity 0.84)
- Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling (CVPR 2024 · similarity 0.84)
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.
