# Visual Attention Network

**Authors:** Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu  
**arXiv:** [2202.09741](https://arxiv.org/abs/2202.09741)  
**Date:** 2022-02-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.09741v5/

## Key points

- Authors address challenges of applying self-attention mechanisms in computer vision due to 2D nature of images
- Neglecting 2D structures of images when treating them as 1D sequences
- High computational complexity for high-resolution images
- Focus on spatial adaptability while overlooking channel adaptability
- Proposed solution: novel linear attention mechanism called large kernel attention (LKA) for self-adaptive and long-range correlations within self-attention
- Introduced neural network architecture based on LKA named Visual Attention Network (VAN)
- VAN outperforms similar-sized vision transformers (ViTs) and convolutional neural networks (CNNs across various computer vision tasks
- VAN-B6 achieves accuracy of 87.8% on ImageNet benchmark and sets new state-of-the-art performance with Panoptic Quality (PQ) score of 58.2 for panoptic segmentation
- VAN-B2 surpasses Swin-T by 4% in mean Intersection over Union (mIoU) for semantic segmentation on ADE20K benchmark and shows a 2.6% improvement in Average Precision (AP) for object detection on COCO dataset
- Visual Attention Network provides groundbreaking method addressing key challenges in applying self-attention mechanisms to computer vision tasks
- Simplicity yet effectiveness establishes VAN as strong baseline model for community to build upon
- Availability of code on GitHub enhances accessibility and reproducibility for researchers interested in exploring this innovative approach to visual attention modeling

## Summary

In their paper titled "Visual Attention Network," authors Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu address the challenges of applying self-attention mechanisms in computer vision due to the 2D nature of images. They highlight three main challenges: neglecting the 2D structures of images when treating them as 1D sequences, the high computational complexity for high-resolution images, and the focus on spatial adaptability while overlooking channel adaptability. To overcome these limitations, the authors propose a novel linear attention mechanism called large kernel attention (LKA) that enables self-adaptive and long-range correlations within self-attention. The authors introduce a neural network architecture based on LKA named Visual Attention Network (VAN). Despite its simplicity, VAN outperforms similar-sized vision transformers (ViTs) and convolutional neural networks (CNNs) across various computer vision tasks such as image classification, object detection, semantic segmentation, panoptic segmentation, and pose estimation. For instance,VAN-B6 achieves an impressive accuracy of 87.8% on the ImageNet benchmark and sets a new state-of-the-art performance with a Panoptic Quality (PQ) score of 58.2 for panoptic segmentation. Additionally,VAN-B2 surpasses Swin-T by 4% in mean Intersection over Union (mIoU) for semantic segmentation on the ADE20K benchmark and shows a 2.6% improvement in Average Precision (AP) for object detection on the COCO dataset. Overall,the Visual Attention Network provides a groundbreaking method that addresses key challenges in applying self-attention mechanisms to computer vision tasks.The simplicity yet effectiveness of VAN establishes it as a strong baseline model for the community to build upon.The availability of code on GitHub further enhances accessibility and reproducibility for researchers interested in exploring this innovative approach to visual attention modeling.