Data augmentation method for domain generalization in object detection

Abstract

For the application of deep learning models in real-world scenarios, it is crucial to consider their robustness across a wide range of domains. Data augmentation aims to improve the robustness of a model to the domain. However, most of these data augmentation techniques have been studied in the field of image classification. When these approaches are applied to object detection, the semantic features of some objects can be damaged, which can lead to imprecise object localization and misclassification. In this paper, an object-aware data augmentation method, which is called OA-Mix, is proposed to address these problems. The method generates multi-domain data using a multi-level transformation and an object-aware mixing strategy. OA-Mix outperforms state-of-the-art methods on the benchmark to evaluate robustness in corrupted domains.