Augmentation, Mixing, and Consistency Regularization for Domain Generalization

This work addresses the prevalent challenge in contemporary deep neural networks (DNNs), in which performance diminishes when confronted with testing data that differ in distribution from the training data. We approach this problem using Domain Generalization (DG), training the model on multiple related source domains to bolster its performance on an unseen target domain. Our approach is a tripartite solution involving data augmentation, mixing of style information, and consistency regularization. Data augmentation is achieved by mixing the amplitude spectrum of two distinct images in the frequency domain. Given that image style is intrinsically tied to the visual domain, the style information is mixed into the lower layers of the neural network. This step familiarizes the model with a range of features, enhancing its ability to generalize on unseen target data. Consistency regularization is then introduced to reduce the prediction error between the original and augmented samples, further elevating the performance. Using three distinct benchmarks, extensive experiments were performed that confirm the Augmentation, Mixing, and Consistency Regularization (AMCR) framework on unseen target domains in comparison with existing state-of-the-art (SOTA) methods. The findings underscore the value of DNNs that can effectively generalize across diverse environments, particularly in real-world applications such as autonomous driving.