Research map / Architectures and efficiency / Attention and transformers
Visformer: The Vision-Friendly Transformer
Zhengsu Chen, Lingxi Xie, Jianwei Niu, Xuefeng Liu, Longhui Wei, Qi Tian
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 global average pooling instead of classification token (ablation) | DeiT-S · ImageNet (base setting) | top-1 accuracy: 64.17 -> 69.44 | Improved |
| add step-wise patch embeddings instead of large patch embedding (ablation) | Net1 · ImageNet (base setting) | top-1 accuracy: 69.81 -> 73.01 | Improved |
| add step-wise patch embeddings instead of large patch embedding (ablation) | Net1 · ImageNet (elite setting) | top-1 accuracy: 80.16 -> 81.35 | Improved |
| add stage-wise design (ablation) | Net2 · ImageNet (base vs elite setting) | top-1 accuracy: base 73.01 -> 75.76; elite 81.35 -> 80.19 | Mixed |
| add batch norm instead of layer norm (ablation) | Net3 · ImageNet (elite setting) | top-1 accuracy: 80.19 -> 80.97 | Improved |
| add 3x3 convolutions in feed-forward blocks (ablation) | Net4 · ImageNet (base vs elite setting) | top-1 accuracy: base 76.49 -> 77.37; elite 80.97 -> 80.15 | Mixed |
| add position embedding (ablation) | DeiT-S · ImageNet (elite setting) | top-1 accuracy: elite performance decreases significantly by 3.95% when removed | Improved |
| add position embedding (ablation) | Net5 · ImageNet (elite setting) | top-1 accuracy: 79.86 -> 80.15 | Improved |
| add self-attention instead of convolution (ablation) | Net6 · ImageNet (elite setting) | top-1 accuracy: 79.01 -> 79.86 | Improved |
| add self-attention in stage 1 instead of bottleneck blocks (ablation) | Net5 · ImageNet (elite setting) | top-1 accuracy: 80.15 -> 80.13 | No clear effect |
| add self-attention in stage 3 instead of bottleneck blocks (ablation) | Net5 · ImageNet (elite setting) | top-1 accuracy: 79.59 -> 80.15 | Improved |
| add bottleneck blocks with 3x3 convolutions in stage 1 only (ablation) | Net4 · ImageNet (elite setting) | top-1 accuracy: 80.97 -> 81.10 | Improved |
Closest papers by meaning
- AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition (NeurIPS 2022 · similarity 0.90)
- FeedFormer: Revisiting Transformer Decoder for Efficient Semantic Segmentation (AAAI 2023 · similarity 0.89)
- A Survey on Vision Transformer (TPAMI 2022 · similarity 0.89)
- Conv2Former: A Simple Transformer-Style ConvNet for Visual Recognition (TPAMI 2024 · similarity 0.87)
- Training Object Detectors From Scratch: An Empirical Study in the Era of Vision Transformer (CVPR 2022 · similarity 0.87)
- Transformer-Based Visual Segmentation: A Survey (TPAMI 2024 · similarity 0.87)
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.
