Research map / Detection and segmentation / CNN features
Vision Transformers: From Semantic Segmentation to Dense Prediction
Li Zhang, Jiachen Lu, Sixiao Zheng, Xinxuan Zhao, Xiatian Zhu, Yanwei Fu, Tao Xiang, Jianfeng Feng, Philip H. S. Torr
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 ImageNet-21K pretraining of the Transformer (vs random init) (ablation) | SETR-PUP T-Large · Cityscapes val | mIoU (%): 42.27 -> 78.39 | Improved |
| add auxiliary loss heads (ablation) | SETR-Naive · Cityscapes val (40k) | mIoU (%): 76.73 -> 77.37 | Improved |
| add auxiliary loss heads (ablation) | SETR-Naive · Cityscapes val (80k) | mIoU (%): 77.23 -> 77.90 | Improved |
| use 4 progressive upsample layers instead of 2 (ablation) | SETR-PUP · Cityscapes val (40k) | mIoU (%): 77.48 -> 78.39 | Improved |
| progressive upsampling decoder (PUP) instead of MLA (ablation) | SETR T-Large · Cityscapes val (40k) | mIoU (%): 76.65 -> 78.39 | Improved |
| MLA decoder instead of PUP (ablation) | SETR T-Large · ADE20K val (MS) | mIoU (%): 50.1 -> 50.3 | No clear effect |
| replace ViT IN-21K pretraining with BEiT IN-21K (ablation) | SETR-PUP T-Large · Cityscapes val (80k) | mIoU (%): 79.34 -> 80.43 | Improved |
| replace ViT IN-21K pretraining with BEiT IN-21K (ablation) | SETR-PUP T-Large · ADE20K val (SS) | mIoU (%): 48.6 -> 52.5 | Improved |
| use DeiT IN-1K pretraining instead of ViT IN-21K on T-Base (ablation) | SETR-PUP T-Base · Cityscapes val (80k) | mIoU (%): 78.02 -> 79.45 | Improved |
| use DeiT IN-1K T-Base instead of ViT IN-1K... T-Large on ADE20K (ablation) | SETR-PUP · ADE20K val (SS) | mIoU (%): 48.6 -> 46.3 | Hurt |
| pretrain FCN on IN-21K instead of IN-1K (ablation) | FCN R-101 · ADE20K val | mIoU (%): 39.91 -> 42.17 | Improved |
| pretrain FCN on IN-21K instead of IN-1K (ablation) | FCN R-101 · Cityscapes val | mIoU (%): 73.93 -> 76.38 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- CCNet — CCNet: Criss-Cross Attention for Semantic Segmentation (ICCV 2019) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Vision Transformers for Dense Prediction (ICCV 2021 · similarity 0.94)
- AiluRus: A Scalable ViT Framework for Dense Prediction (NeurIPS 2023 · similarity 0.93)
- A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation (ECCV 2022 · similarity 0.91)
- Locality-Attending Vision Transformer (ICLR 2026 · similarity 0.91)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.91)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.90)
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.
