Understanding Convolution for Semantic Segmentation
1 Jun 2018
thought
1、增加的可学习参数形式上不一定对应，能变换回来就行
2、直觉上的不合理要仔细分析
motivation
1、上采样更好地恢复信息，bilinear upsampling is not learnable, deconvolution zeros have to be padded
2、空洞卷积有 gridding 现象
solution
dense upsampling convolution (DUC)
增加通道数，将conv恢复的feature加在channel维度上，再reshape到original resolution
论文效果比 deconv 略好
hybrid dilated convolution (HDC)
级联不同 rate 的 dilated conv，覆盖掉 grid
参考
https://blog.csdn.net/u011974639/article/details/79460893
https://www.cnblogs.com/ansang/p/9003513.html
https://zhuanlan.zhihu.com/p/26659914
