Distribution-Specific Augmentation for Domain Generalization

The generalization capability of deep neural networks usually deteriorates when the data changes distribution during test time. One type of such distribution shift is domain shift, where different domains are responsible for changing the underlying distribution of the associated data. This study proposes distribution-specific augmentation to mitigate domain generalization in multi-class image classification. Our proposed technique focuses more on sampling data related to specific classes. We do this assuming that class-related features are less likely to change significantly. To achieve this, we employ a background modeling technique called Robust Principal Component Analysis—Principal Component Pursuit (RPCA-PCP), which separates features specific to each class from the overall data. We assess the method on a simulated dataset and the iWildCam 2020 dataset. While the proposed method is applied successfully to the synthetic dataset, it only marginally improves overall test accuracy performance on the iWildCam dataset. However, the method highlights the dichotomy between class-relevant and irrelevant distributions in the data, which could inspire future research building on the ideas presented.