Distance-Based Class Activation Map for Metric Learning

Yeqing Shen,Huimin Ma,Xiaowen Zhang,Tianyu Hu,Yuhan Dong
DOI: https://doi.org/10.1007/978-3-030-88013-2_28
2021-01-01
Abstract:The interpretability of deep neural networks can serve as reliable guidance for algorithm improvement. By visualizing class-relevant features in the form of heatmap, the Class Activation Map (CAM) and derivative versions have been widely exploited to study the interpretability of softmax-based neural networks. However, CAM cannot be adopted directly for metric learning, because there is no fully-connected layer in metric-learning-based methods. To solve this problem, we propose a Distance-based Class Activation Map (Dist-CAM) in this paper, which can be applied to metric learning directly. Comprehensive experiments are conducted with several convolutional neural networks trained on the ILSVRC 2012 and the result shows that Dist-CAM can achieve better performance than the original CAM in weakly-supervised localization tasks, which means the heatmap generated by Dist-CAM can effectively visualize class-relevant features. Finally, the applications of Dist-CAM on specific tasks, i.e., few-shot learning, image retrieval and re-identification, based on metric learning are presented.
What problem does this paper attempt to address?