Research map / Architectures and efficiency / Attention and transformers
Vision Transformers for Dense Prediction
René Ranftl, Alexey Bochkovskiy, Vladlen Koltun
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 |
|---|---|---|---|
| use ViT-Hybrid backbone instead of ResNet50 (ablation) | DPT · Depth reduced validation set | mean relative deviation: 0.0935 -> 0.0783 | Improved |
| use ViT-Large instead of ViT-Base (ablation) | DPT · Depth reduced validation set | mean relative deviation: 0.0819 -> 0.0778 | Improved |
| use ViT-Hybrid instead of ViT-Base (ablation) | DPT · Depth reduced validation set | mean relative deviation: 0.0819 -> 0.0783 | Improved |
| use distilled DeIT-Base-Dist pretraining instead of ViT-Base (ablation) | DPT · Depth reduced validation set | mean relative deviation: 0.0819 -> 0.0796 | Improved |
| tap ResNet low-level stages R0,R1 instead of transformer layers (ablation) | DPT-Hybrid · Depth reduced validation set | mean relative deviation: 0.0787 -> 0.0733 | Improved |
| tap shallow and deep layers {3,6,9,12} instead of {9,10,11,12} (ablation) | DPT-Base · Depth reduced validation set | mean relative deviation: 0.0828 -> 0.0822 | Improved |
| project the readout token instead of ignoring it (ablation) | DPT-Base · Depth reduced validation set | mean relative deviation: 0.0822 -> 0.0819 | Improved |
| add the readout token to the tokens instead of ignoring it (ablation) | DPT-Base · Depth reduced validation set | mean relative deviation: 0.0822 -> 0.0831 | Hurt |
| train on larger MIX 6 dataset with the MiDaS convolutional network | MiDaS · Zero-shot depth | DIW WHDR: 12.46 -> 12.95 | Improved |
| replace MiDaS conv network with DPT-Large on MIX 6 | MiDaS · Zero-shot depth | DIW WHDR: 12.95 -> 10.82 | Improved |
| use DPT-Large instead of DPT-Hybrid for ADE20K | DPT · ADE20K val | mIoU: 49.02 -> 47.63 | Hurt |
| use DPT-Hybrid for monocular depth zero-shot on TUM instead of MiDaS | MiDaS · TUM (zero-shot) | delta>1.25 error: 12.51 -> 10.89 | Improved |
Closest papers by meaning
- Vision Transformers: From Semantic Segmentation to Dense Prediction (IJCV 2024 · similarity 0.94)
- Expediting Large-Scale Vision Transformer for Dense Prediction without Fine-tuning (NeurIPS 2022 · similarity 0.92)
- Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction Without Convolutions (ICCV 2021 · similarity 0.90)
- AiluRus: A Scalable ViT Framework for Dense Prediction (NeurIPS 2023 · similarity 0.90)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.90)
- HRFormer: High-Resolution Vision Transformer for Dense Predict (NeurIPS 2021 · 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.
