Research map / Detection and segmentation / CNN features
Sparse Refinement for Efficient High-Resolution Semantic Segmentation
Zhijian Liu, Zhuoyang Zhang, Samir Khaki, Shang Yang, Haotian Tang, Chenfeng Xu, Kurt Keutzer, Song Han
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 sparse high-resolution refinement to dense low-res prediction | SegFormer-B5 · Cityscapes val | latency (ms): 140.6 -> 38.5 (3.7x), mIoU 81.1 -> 81.2 | Improved |
| add SparseRefine to SegNeXt-L | SegNeXt-L · Cityscapes val | mean IoU: 83.0 -> 82.8 | Hurt |
| add SparseRefine to HRNet-W48 vs dense half-resolution | HRNet-W48 · Cityscapes val | mean IoU: 79.2 -> 80.9 | Improved |
| use SparseRefine instead of PointRend refinement | HRNet-W48 · Cityscapes val | mean IoU: 79.9 -> 80.9 | Improved |
| use sparse pixel refinement instead of patch refinement | HRNet-W48 · Cityscapes val | latency (ms): 44.1 -> 32.4 | Improved |
| use entropy selector instead of random selector (ablation) | SparseRefine HRNet-W48 · Cityscapes val | mIoU: 79.2 -> 80.9 | Improved |
| use entropy selector instead of magnitude selector (ablation) | SparseRefine HRNet-W48 · Cityscapes val | mIoU: 80.2 -> 80.9 | Improved |
| use entropy selector instead of learnable selector (ablation) | SparseRefine HRNet-W48 · Cityscapes val | selector latency: 4.0ms -> 2.3ms | Improved |
| add gated ensembler instead of entropy ensembler (ablation) | SparseRefine HRNet-W48 · Cityscapes val | mIoU: 80.3 -> 80.9 | Improved |
| add logits or features as input to sparse extractor (ablation) | SparseRefine HRNet-W48 · Cityscapes val | mIoU: 80.9 -> 80.2 | Hurt |
| refine all pixels densely instead of sparse 11.8% (ablation) | SparseRefine HRNet-W48 · Cityscapes val | mIoU: 80.9 -> 80.5 | Hurt |
| lower entropy threshold alpha from 0.8 to 0.1 (more pixels refined) (ablation) | SparseRefine HRNet-W48 · Cityscapes val | mIoU: 79.9 -> 81.1 (latency 24.6 ms -> 40.1 ms) | Improved |
Closest papers by meaning
- FineRS: Fine-grained Reasoning and Segmentation of Small Objects with Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Stronger Fewer & Superior: Harnessing Vision Foundation Models for Domain Generalized Semantic Segmentation (CVPR 2024 · similarity 0.86)
- Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 2024 · similarity 0.86)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.86)
- Sparse All-Layer Connector for Domain Generalised Semantic Segmentation (NeurIPS 2026 · similarity 0.86)
- Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2026 · 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.
