Title

Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks
代码：https://github.com/hujie-frank/GENet

这篇文章的作者hujie也是SENet的作者，算是channel attention的开山之作，发表在CVPR2018。GENet算是SENet后来的发展。这儿我把SENet文章中的图也贴一下，当作复习。

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图1

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图2[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图3

Summary

作者在SENet的基础上，进一步探讨什么样的gather + excite组合更能提升效果。结果表明：gather操作采用更大的感受野更好，excite操作增加1*1卷积更好。

Method(s)

其实作者一共比较了三种结构，分别将其定义为：[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图4

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图5：完全无参数，采用pooling + resize(neatest interpolation) + sigmoid + multiply

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图6

pooling的时候可以采用不同的压缩比例，对比结果表明global avg pooling效果是最好的，即下图右优于下图左

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图7[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图8

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图9：用depth-wise conv做压缩，excite操作和[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图10一样，也是右边的效果好

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图11[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图12

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图13：在[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图14 基础上，在excite部分增加了1*1卷积，这块儿paper没说得很清楚

这三种结构的比较结果：[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图15效果最好

[GENet]Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks - 图16

Notes

我目前的模型，时延符合要求了，但一段时间后会降帧，说明功耗高，需要降op。打算沿着这个思路测试和优化。