Research map / Architectures and efficiency / Attention and transformers
Demystify Transformers & Convolutions in Modern Image Deep Networks
Xiaowei Hu, Min Yong Shi, Wei‐Yun Wang, Sitong Wu, Linjie Xing, Wenhai Wang, Xizhou Zhu, Lewei Lu, Jie Zhou, Xiaogang Wang, Yu Qiao, Jifeng Dai
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 InternImage deformable convolution instead of PVT spatial-reduction attention | unified architecture (Tiny) · ImageNet-A | accuracy: 26.1 -> 31.9 | Improved |
| use deformable convolution v3 instead of depth-wise convolution | unified architecture (Tiny) · ImageNet-A | accuracy: 23.3 -> 31.9 | Improved |
| use halo attention instead of PVT spatial-reduction attention | unified architecture (Tiny) · ImageNet-C | 1-mCE: 56.1 -> 57.9 | Improved |
| use PVT global attention instead of InternImage DCNv3 | unified architecture (Tiny) · FGSM eps=8/255 | adversarial accuracy: 41.2 -> 41.7 | No clear effect |
| use Swin shifted-window attention instead of halo attention | unified architecture (Tiny) · FGSM eps=8/255 | adversarial accuracy: 38 -> 32.7 | Hurt |
| use deformable convolution instead of PVT for translation invariance at 64 px | unified architecture (Tiny) · ImageNet-1k translation | consistency: 87.6 -> 89.2 | Improved |
| use deformable convolution instead of depth-wise convolution for rotation 45 degrees | unified architecture (Tiny) · ImageNet-1k rotation | accuracy: 75.7 -> 78.3 | Improved |
| use deformable convolution instead of Swin for scale 0.25 on COCO | unified architecture · COCO | box mAP: 23.3 -> 25.2 | Improved |
| scale up model from Micro to Tiny | U-InternImage · ImageNet-1k | clean accuracy: 75.3 -> 83.3 | Improved |
Closest papers by meaning
- MetaFormer Is Actually What You Need for Vision (CVPR 2022 · similarity 0.87)
- Dynamic Spatial Sparsification for Efficient Vision Transformers and Convolutional Neural Networks (TPAMI 2023 · similarity 0.87)
- Do Vision Transformers See Like Convolutional Neural Networks? (NeurIPS 2021 · similarity 0.87)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.86)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.86)
- LSNet: See Large, Focus Small (CVPR 2025 · 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.
