论文笔记——Factorized Convolutional Neural Networks

时间:2023-03-09 16:14:17
论文笔记——Factorized Convolutional Neural Networks

1. 论文思想

将3D卷积分解为spatial convolution in each channel and linear projection across channels.
(spatial convolution + linear projection.)

2. 两种卷积对比

论文笔记——Factorized Convolutional Neural Networks

论文笔记——Factorized Convolutional Neural Networks

3. 总结

简单概括就是spatial conv + linear projection,但是在spatial conv的时候用了一个residual connection,感觉很有道理,例如是一个vertical edge detector,那么horizontal information将丢失。这个和后来的MobileNet中的depthwise conv + pointwise conv非常的像。