Deep Relevant Feature Focusing for Out-of-Distribution Generalization

Convolution Neural Networks (CNNs) often fail to maintain their performance when they confront new test domains. Unlike human's strong ability of abstraction and connection, CNNs learn everything relevant and irrelevant from their training data while humans can understand its essential features and form. In this paper, we propose a method to improve the cross-domain object recognition ability from the model feature level: Our method masks the partial values of the feature maps to force models to focus on potentially important features. Multiple experiments on the PACS and VLCS confirm our intuition and show that this simple method outperforms previous domain generalization solutions.