论文学习CAM系列

Learning Deep Features for Discriminative Localization

Research Background

1. In CNNs, the convolution units perform as an object detector. However, the fully-connected layers wipe off this remarkable localization ability.
2. The author proposed the Global Average Pooling layer to act as a structural regularizer. In this paper, the authors proposed that the global average pooling layer can help the network to retain the localization ability with some tweaking.

Proposed method

Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization

Summary

This paper proposed an approach named Gradient-weighted Class Activation Mapping. The method makes the gradient of the target flow into the final convolutional layer to produce a coarse localization map. This localization map can highlight the important regions in the image for predicting the concept.