Research map / Architectures and efficiency / Attention and transformers
Lightweight Vision Transformer with Bidirectional Interaction
Qihang Fan, Huaibo Huang, Xiaoqiang Zhou, Ran He
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 |
|---|---|---|---|
| replace add+linear local-global fusion with bidirectional adaptive interaction (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 76.2 -> 77.6 | Improved |
| replace cat+linear fusion with bidirectional interaction (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 76.6 -> 77.6 | Improved |
| replace element-wise multiplication fusion with bidirectional interaction (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.1 -> 77.6 | Improved |
| replace pooling downsampling with fine-grained downsampling (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.2 -> 77.6 | Improved |
| replace overlapping large-stride conv with fine-grained downsampling (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.3 -> 77.6 | Improved |
| add conditional positional encoding (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.4 -> 77.6 | Improved |
| add local adaptive aggregation and bidirectional interaction to global-only (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 76.2 -> 77.6 | Improved |
| add convolution stem (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.2 -> 77.6 | Improved |
| add extra sigmoid activation in local aggregation (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.6 -> 77.6 | No clear effect |
| increase DWConv kernel from 3 to staged 3,5,7,7 (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.3 -> 77.6 | Improved |
| use uniform large kernel 9 in all stages vs staged kernels (ablation) | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 77.6 -> 77.4 | Hurt |
| replace Max-SA with FASA on same backbone | FAT-B0 · ImageNet-1K | Top-1 accuracy (%): 75.7 -> 77.6 | Improved |
Closest papers by meaning
- DaViT: Dual Attention Vision Transformers (ECCV 2022 · similarity 0.87)
- Fcaformer: Forward Cross Attention in Hybrid Vision Transformer (ICCV 2023 · similarity 0.87)
- Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · similarity 0.86)
- Efficient High-Order Spatial Interactions for Visual Perception (TPAMI 2025 · similarity 0.86)
- CARE Transformer: Mobile-Friendly Linear Visual Transformer via Decoupled Dual Interaction (CVPR 2025 · similarity 0.86)
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · 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.
