Learning Deep Features for Discriminative Localization

# Learning Deep Features for Discriminative Localization

# 作者：Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, Antonio Torralba

# 单位：MIT

# 发表：CVPR 2016

# 摘要

# 阅读

# 论文的目的及结论

# 论文的实验

# 论文的方法

核心代码非常简单，提取到特征图和目标类别全连接的权重，直接加权求和，再经过relu操作去除负值，最后归一化获取CAM，具体如下:

# 获取全连接层的权重
self._fc_weights = self.model._modules.get(fc_layer).weight.data
# 获取目标类别的权重作为特征权重
weights=self._fc_weights[class_idx, :]
# 这里self.hook_a为最后一层特征图的输出
batch_cams = (weights.unsqueeze(-1).unsqueeze(-1) * self.hook_a.squeeze(0)).sum(dim=0)
# relu操作,去除负值
batch_cams = F.relu(batch_cams, inplace=True)
# 归一化操作
batch_cams = self._normalize(batch_cams)

1
2
3
4
5
6
7
8
9
10

# 论文的背景

# 总结

# 论文的贡献

利用 GAP 获取 CAM 的开山之作

# 论文的不足

# 论文如何讲故事

# 参考资料

https://arxiv.org/abs/1512.04150
https://github.com/zhoubolei/CAM
- https://github.com/zhoubolei/CAM/blob/master/pytorch_CAM.py
https://cloud.tencent.com/developer/article/1674200

上次更新: 2021/11/03, 23:35:28

← Weakly Supervised Learning of Instance Segmentation with Inter-pixel Relations Convolutional Random Walk Networks for Semantic Image Segmentation→