DANet：Dual Attention Network for Scene Segmentation
Overview
![[论文学习] DANet：Dual Attention Network for Scene Segmentation](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzMxOS9jMGY3NWNhNWNhNDE4NWY4ODlkNTI5NWQyOThhN2M4Zi5wbmc= "[论文学习] DANet：Dual Attention Network for Scene Segmentation")
基础网络为dilated ResNet（与DeepLab相同），最后得到的feature map大小为输入图像的1/8.之后使两个并行的attention module分别捕获spatial 和 channel 的依赖性，最后整合两个attention module的输出得到更好的特征表达。
Position Attention Module
![[论文学习] DANet：Dual Attention Network for Scene Segmentation](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzcxMy8xYTdmZGRiY2M0MTc1MjFlMDM1YjhlMmUwOTgyYjExMS5wbmc= "[论文学习] DANet：Dual Attention Network for Scene Segmentation")
Channel Attention Module
![[论文学习] DANet：Dual Attention Network for Scene Segmentation](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzIxOC9iYzE2YWRmYzI3ZWQ1YTRmZmUzNzNiNjkxOGQyZTE5Mi5wbmc= "[论文学习] DANet：Dual Attention Network for Scene Segmentation")
Result
![[论文学习] DANet：Dual Attention Network for Scene Segmentation](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzIyNC8zMmM5ZWQ5YWQ4M2IxOGZmOTI1Y2UzNGRkZmFlM2U0OC5wbmc= "[论文学习] DANet：Dual Attention Network for Scene Segmentation")
![[论文学习] DANet：Dual Attention Network for Scene Segmentation](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzMwNi8wNzg1YTdhNmNlYTFmYzIxMDk4YzlmMWFmOGNmNjAxYS5wbmc= "[论文学习] DANet：Dual Attention Network for Scene Segmentation")
总结
![[论文学习] DANet：Dual Attention Network for Scene Segmentation](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzk0NS84MmU1ZjdmNDEwZmFhN2U2OTZlZWNmNzlmNzRhYzg5MS5wbmc= "[论文学习] DANet：Dual Attention Network for Scene Segmentation")
