记录一下看过的资料,主要以DeepLab 相关
DeepLab v1 + v2
https://blog.csdn.net/zhuzemin45/article/details/79769154
https://blog.csdn.net/Dlyldxwl/article/details/81148810
v1 中涉及到的 FCN
https://blog.csdn.net/zhuzemin45/article/details/79647862
https://blog.csdn.net/P_LarT/article/details/83904661#FCN_30(这个对FCN的解释老详细老有意思了,有空看一下)
deconv 概念 (upSampling):
https://blog.csdn.net/xiaojiajia007/article/details/75041651
https://blog.csdn.net/itleaks/article/details/80336825 (要两个结合起来看)
几种有意思的upSampling 方法:
https://blog.csdn.net/u014451076/article/details/79156967 没有看到这篇博文我都以为FCN的upsampling方法是 deconv,但是看代码里好像不太对,原来是用较简单粗暴又快速的双线性插值
v1中涉及到的 fully CRF
https://www.jiqizhixin.com/articles/2018-05-23-3
https://kexue.fm/archives/4695
https://zhuanlan.zhihu.com/p/28465510 (详细讲解,未看)
v2 中涉及到的感受野概念(我竟被这样一个小知识点给hit到...
https://zhuanlan.zhihu.com/p/31004121
v2中涉及到的ResNet 101 概念
https://blog.csdn.net/briblue/article/details/83544381
https://zhuanlan.zhihu.com/p/33613512 (坡的蒋老师的讲解,挺清晰的,专栏中有专门的代码实现)
https://www.cnblogs.com/wzyuan/p/9880342.html(ResNet pytorch 官方代码讲解)
关于ResNet 残差证明部分:https://blog.csdn.net/wspba/article/details/60750007
Residual Bottleneck:https://blog.csdn.net/u011304078/article/details/80683985
关于dilation 和 dilated convolution:
https://blog.csdn.net/wangyuxi__/article/details/83003357
https://www.zhihu.com/question/54149221
结果顺便看了v3的知识
https://blog.csdn.net/qq_21997625/article/details/87080576
https://blog.csdn.net/u011974639/article/details/79144773
关于pytorch中导入预训练模型流程:
https://blog.csdn.net/u014380165/article/details/79119664
pytorch 与交叉熵相关的ignore-index在语义分割中作用的介绍
https://discuss.pytorch.org/t/when-to-use-ignore-index/5935
https://discuss.pytorch.org/t/ignore-index-in-nn-crossentropyloss-for-semantic-segmentation/17645 (未看完
pytorch关于语义分割数据集建立的详细解释
https://blog.csdn.net/Teeyohuang/article/details/82108203
关于语义分割问题上的准确率计算方法 IoU 以及 神奇的fical loss:
https://tianws.github.io/skill/2018/10/30/miou/
https://www.jianshu.com/p/ab0a10c2e710
几种分割算法的loss
https://blog.csdn.net/wangdongwei0/article/details/84576044