Research map / Detection and segmentation / CNN features
Low-Resolution Self-Attention for Semantic Segmentation
Yu-Huan Wu, Shichen Zhang, Yun Liu, Le Zhang, Xin Zhan, Daquan Zhou, Jiashi Feng, Ming‐Ming Cheng, Liangli Zhen
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 |
|---|---|---|---|
| increase fixed pooled size of LRSA from 4x4 to 16x16 (ablation) | LRFormer-S · ADE20K val | mIoU: 46.3% -> 48.5% | Improved |
| increase pooled size beyond 16x16 (32x32, 48x48, 64x64) (ablation) | LRFormer-S · ADE20K val | mIoU: 48.5% -> 48.6%, 48.7%, 48.5% with FLOPs 40G -> 52G, 74G, 108G | No clear effect |
| add depth-wise convolution before LRSA (ablation) | LRFormer-S · ADE20K val | mIoU: 48.0% -> 48.5% | Improved |
| add depth-wise convolution in the FFN (ablation) | LRFormer-S · ADE20K val | mIoU: 47.1% -> 48.5% | Improved |
| add both depth-wise convolutions (ablation) | LRFormer-S · ADE20K val | mIoU: 44.7% -> 48.5% | Improved |
| add both depth-wise convolutions (ablation) | LRFormer-S · ImageNet-1K (100 epochs) | top-1 accuracy: 80.4% -> 81.6% | Improved |
| use 16x16 pooled size instead of 4x4 on small objects (ablation) | LRFormer-S · ADE20K small-object classes | mIoU: 33.7% -> 36.2% | Improved |
| use LRSA decoder instead of OCR head (ablation) | LRFormer-S · ADE20K val | mIoU: 48.0% -> 49.5% | Improved |
| use LRSA decoder instead of PPM, DA or CC heads (ablation) | LRFormer-S · ADE20K val | mIoU: 48.4% -> 49.5% (PPM), 48.9% -> 49.5% (DA), 48.6% -> 49.5% (CC) | Improved |
| replace SegFormer-B2 decoder with LRSA decoder (ablation) | SegFormer-B2 · ADE20K val | mIoU: 46.5% -> 47.3% | Improved |
| increase decoder dimension from 128 to 384 (ablation) | LRFormer-S · ADE20K val | mIoU: 47.8% -> 49.5% | Improved |
| increase decoder dimension beyond 384 (512, 768, 1024) (ablation) | LRFormer-S · ADE20K val | mIoU: 49.5% -> 49.6%, 49.2%, 49.2% | No clear effect |
Closest papers by meaning
- Rethinking Semantic Segmentation From a Sequence-to-Sequence Perspective With Transformers (CVPR 2021 · similarity 0.89)
- Embedding-Free Transformer with Inference Spatial Reduction for Efficient Semantic Segmentation (ECCV 2024 · similarity 0.88)
- Rethinking Self-Supervised Semantic Segmentation: Achieving End-to-End Segmentation (TPAMI 2024 · similarity 0.87)
- Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 2024 · similarity 0.87)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.86)
- Multi-Scale Representations by Varying Window Attention for Semantic Segmentation (ICLR 2024 · 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.
