Research map / Architectures and efficiency / Attention and transformers
Fcaformer: Forward Cross Attention in Hybrid Vision Transformer
Haokui Zhang, Wenze Hu, Xiaoyu Wang
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 naive forward cross attention without scale factors (ablation) | ConvViT (hybrid, global attention) · ImageNet-1K val | top-1 accuracy (%): 79.4 -> 79.4 | No clear effect |
| add learnable scale factors (LSFs) for cross tokens (ablation) | FcaFormer · ImageNet-1K val | top-1 accuracy (%): 79.4 -> 79.9 | Improved |
| add token merge and enhancement (TME) module (ablation) | FcaFormer · ImageNet-1K val | top-1 accuracy (%): 79.9 -> 80.3 | Improved |
| add early convolution stages (ablation) | Swin-1G · ImageNet-1K val | top-1 accuracy (%): 78.4 -> 78.8 | Improved |
| replace window attention with global attention (ablation) | ConvSwin · ImageNet-1K val | top-1 accuracy (%) and latency: 78.8 -> 79.4; latency 269 -> 300 ms | Mixed |
| replace DeiT-T with forward cross attention (FcaFormer-L0) | DeiT-Tiny · ImageNet-1K val | top-1 accuracy (%): 72.2 -> 74.3 | Improved |
| use FcaFormer-L2 instead of ConvNext-Tiny | hybrid backbone · ImageNet-1K val | top-1 accuracy (%): 82.1 -> 83.1 | Improved |
| use FcaFormer-L2 backbone in UperNet instead of ConvNext-T | UperNet · ADE20K | mIoU: 46.7 -> 47.6 | Improved |
| use FcaFormer-L2 backbone in Mask-RCNN instead of ConvNext-T | Mask-RCNN 3x · COCO 2017 | box AP: 46.2 -> 47.0 | Improved |
| add learnable scale factors to calibrate cross tokens (attention analysis) (ablation) | FcaFormer · attention score maps | attention to cross tokens | Improved |
Closest papers by meaning
- CrossFormer: A Versatile Vision Transformer Hinging on Cross-scale Attention (ICLR 2022 · similarity 0.89)
- CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification (ICCV 2021 · similarity 0.88)
- AMixer: Adaptive Weight Mixing for Self-Attention Free Vision Transformers (ECCV 2022 · similarity 0.88)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.87)
- Enhancing Vision Transformer: Amplifying Non-Linearity in Feedforward Network Module (ICML 2024 · similarity 0.86)
- Token Statistics Transformer: Linear-Time Attention via Variational Rate Reduction (ICLR 2025 · similarity 0.86)
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.
