CAM:https://arxiv.org/abs/1512.04150(Bolei Zhou, MIT)
Network in network: https://arxiv.org/abs/1312.4400(Shuicheng Yan, NUS)
CAM:class activation mapping
一句话总结:在原图上显示出对分类的贡献分布
1. CAM总体示意图:GAP然后加权求和(GAP:Global average pooling)
2.CAM公式推导:加权求和交换顺序
3.CAM显示定位功能:上采样回到原图大小
4.对比试验
构建网络:GoogLeNet-GAP
(GoogLeNet结构表)
分类性能比较:
定位性能比较:
bbox策略:
(1)simple:
(2)heuristics:
5.补充上面提到的弱监督定位方法
(1)GoogLeNet-GMP:https://www.di.ens.fr/~josef/publications/Oquab15.pdf
(2)Back Prop:https://arxiv.org/pdf/1312.6034.pdf
(3)NIN:https://arxiv.org/abs/1312.4400