DUC HDC 筆記

Understanding Convolution for Semantic Segmentation

1 Jun 2018

thought

1、增加的可學習參數形式上不一定對應，能變換回來就行
2、直覺上的不合理要仔細分析

motivation

1、上採樣更好地恢復信息，bilinear upsampling is not learnable, deconvolution zeros have to be padded
2、空洞卷積有 gridding 現象

solution

dense upsampling convolution (DUC)
增加通道數，將conv恢復的feature加在channel維度上，再reshape到original resolution
論文效果比 deconv 略好
hybrid dilated convolution (HDC)
級聯不同 rate 的 dilated conv，覆蓋掉 grid

參考
https://blog.csdn.net/u011974639/article/details/79460893
https://www.cnblogs.com/ansang/p/9003513.html
https://zhuanlan.zhihu.com/p/26659914