MUST Augment: Efficient Augmentation with Multi-stage Stochastic Strategy

Data augmentation has been widely used for enhancing data diversity and deep-learning model generalization. Recent research achieved better accuracies on image classification tasks by introducing automated search for optimal augmentation policies, but incurred high computation cost and long search time because of large search spaces and complex searching algorithms. In this paper, we present an augmentation method called MUST (MUlti-Stage sTochastic) Augment which completely skips policy searching. Instead of searching, our method applies a multi-stage augmentation strategy on top of a simple stochastic augmentation mechanism. This multi-stage complexity driven augmentation strategy ensures the whole training process converges smoothly to a good quality model; and within individual stages, it applies augmentation in a stochastic manner and provides both scalability and diversity by introducing more augmentation operations without extra search cost. Our extensive experiments with state-of-the-art results show that our method has advantages in both accuracy and efficiency compared to search-based augmentation methods. Besides image classification, we also examine the general validity of MUST on Face Recognition and Text Detection tasks, and demonstrate the effectiveness of our method across various CV tasks.