2018年发表论文阅读：Convolutional Simplex Projection Network for Weakly Supervised Semantic Segmentation

记笔记目的：刻意地、有意地整理其思路，综合对比，以求借鉴。他山之石，可以攻玉。

《Convolutional Simplex Projection Network for Weakly Supervised Semantic Segmentation》-20180724，一篇来自德国波恩大学与锡根大学的paper。

论文code：

https://github.com/briqr/CSPN

Abstract

The method introduces a novel layer which applies simplex projection on the output of a neural network using area constraints of class objects.
该方法提出了一种新颖的层，该层使用类目标对象的区域约束将单一投影应用于神经网络的输出。

该方法可以自然无缝地与任意CNN架构融合在一起，同时，作者所提出的投影层允许强监督模型通过替换ground truth标签而毫不费力地适应弱监督模型。

1 Introduction：

The task of semantic image segmentation, which requires solving the problem of assigning a semantic class label to each pixel in a given image。这句话极好，可以借鉴！

本文提出的方法更加实用。它将约束直接融入网络层，形成新的网络层，该新网络层可以方便地加入进任何卷积神经网络里面去。

2 Related work：
read history。

3 Convolutional Simplex Projection Network (CSPN)：

还是英语顺眼啊。挑拣关键字眼梳理一下这一小节：

3.1 Simplex Projection Layer

3.2 CSPN for Weakly Supervised Semantic Segmentation

Figure 1 gives an overview of how the simplex projection layer can be applied in a weakly
supervised setting, in which only image-level labels are available. In order to enforce some
given constraints at the last layer, we introduce a softmax layer after the last fully convolu-
tional layer in the network, which performs:

3.3 Simplex Projection Layer in SEC