论文笔记:Learning Deep Features for
旧文赏析 CVPR2016 MIT周博磊
idea很简单,但是很work
对model中判别性区域进行定位(use CAM)
achieve 37.1% top-5 error for object localization on ILSVRC 2014, which is remarkably close to the 34.2% top-5 error achieved by a fully supervised CNN approach.
卷积层能够对物体进行定位,但是FC层丢失了位置信息。replace IT with GAP(global average pooling)
related work:
global average pooling (出自 NIN)
Weakly-supervised object localization:
缺点:not trained end-to-end and require multiple forward passes of a network to localize objects
Class Activation Mapping:
Mc(x,y)直接表示空间位置(x, y)处的activation对于类别c的重要性。
实现起来的话,就是算了加权和。把softmax层和conv5的算了加权 ,得到一个加权的featuremap,然后再把它resize
为什么可以这么做?(公式推导)
以自己的图片为例:label为PDR预测为PDR
可以用在可视化、弱监督学习等等领域
GAP vs GMP
GAP encourages the network to identify the extent of the object
GMP identify just one discriminative part