论文笔记：CVPR Regional Semantic Contrast and Aggregation for Weakly Supervised Semantic Segmentation

任务：image-level标签的弱监督分割

Motivation：numerous semantically similar but visually different instances; for any two different concepts (e.g., cat and dog), all their instances are semantically different, even though some may look very similar with each other。且现有的方法忽略了图像之间的相互关系。

本文提出了regional semantic contrast and aggregation (RCA)。

Method：

A.Regional Semantic Contrast and Aggregation（PCA）：五个模块组成

1/Pseudo-Region Representation

2/Pseudo-Region Memory Bank

3/Regional Semantic Contrast (RSC)

原本的cross-Image

但因为没有准确的标注，利用mixup得到混合样本标注。

利用不同的类别

4/Regional Semantic Aggregation (RSA)

针对数据集，而非单张图。但直接引入memory的原型计算需要很大的计算量，本文先采用k-means聚类，再语义信息聚集。

5/Class Activation Map Prediction

^F经过CAM得到输出

B.Detailed Network Architecture