Research map / Architectures and efficiency / Attention and transformers
Convolutional Embedding Makes Hierarchical Vision Transformer Stronger
Cong Wang, Hongmin Xu, Xiong Zhang, Li Wang, Zhitong Zheng, Haifeng Liu
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 LEWin convolutional projection attention (vs shifted window attention) (ablation) | CETNet-T · ImageNet-1K val (ablation, 224) | top-1 accuracy (%): 82.5 -> 82.7 | Improved |
| use deep-narrow instead of shallow-wide design (ablation) | CETNet-T · ImageNet-1K val (ablation, 224) | top-1 accuracy (%): 82.5 -> 82.7 | Improved |
| use CE instead of Swin patch embedding and merging (ablation) | CETNet shallow-wide · ImageNet-1K val (ablation, 224) | top-1 accuracy (%): 81.5 -> 82.5 | Improved |
| use CE instead of CvT convolutional token embedding (ablation) | CETNet shallow-wide · ImageNet-1K val (ablation, 224) | top-1 accuracy (%): 82.0 -> 82.5 | Improved |
| increase CE depth from 1 to 7 MBConv layers (ablation) | Swin-T · ImageNet-1K val (CE layers, Swin-T based) | top-1 accuracy (%): 81.8 -> 82.6 | Improved |
| increase CE depth from 5 to 7 layers (ablation) | Swin-T · ImageNet-1K val (CE layers, Swin-T based) | top-1 accuracy (%): 82.4 -> 82.6 | No clear effect |
| add convolutional embedding to CvT-13 (ablation) | CvT-13 · ImageNet-1K val | top-1 accuracy (%): 81.6 -> 82.1 | Improved |
| add convolutional embedding to PVT-S (ablation) | PVT-S · ImageNet-1K val | top-1 accuracy (%): 79.8 -> 81.1 | Improved |
| add convolutional embedding to Swin-T (ablation) | Swin-T · ImageNet-1K val | top-1 accuracy (%): 81.3 -> 82.5 | Improved |
| add convolutional embedding to CSWin-T (ablation) | CSWin-T · ImageNet-1K val | top-1 accuracy (%): 82.7 -> 83.6 | Improved |
| replace Fused-MBConv with MBConv in early stage (ablation) | CETNet-T · ImageNet-1K val (ablation, 224) | top-1 accuracy (%): 82.7 -> 82.6 | No clear effect |
| use ResNet blocks instead of Fused-MBConv for CE (ablation) | CETNet-T · ImageNet-1K val (ablation, 224) | top-1 accuracy (%): 82.7 -> 82.0 | Hurt |
Closest papers by meaning
- Incorporating Convolution Designs Into Visual Transformers (ICCV 2021 · similarity 0.91)
- Bridging the Gap Between Vision Transformers and Convolutional Neural Networks on Small Datasets (NeurIPS 2022 · similarity 0.91)
- Do Vision Transformers See Like Convolutional Neural Networks? (NeurIPS 2021 · similarity 0.90)
- The Principle of Diversity: Training Stronger Vision Transformers Calls for Reducing All Levels of Redundancy (CVPR 2022 · similarity 0.90)
- CvT: Introducing Convolutions to Vision Transformers (ICCV 2021 · similarity 0.90)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.89)
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.
