SuperMix: Supervising the Mixing Data Augmentation

In this paper, we propose a supervised mixing augmentation method, termed SuperMix, which exploits the knowledge of a teacher to mix images based on their salient regions. SuperMix optimizes a mixing objective that considers: i) forcing the class of input images to appear in the mixed image, ii) preserving the local structure of images, and iii) reducing the risk of suppressing important features. To make the mixing suitable for large-scale applications, we develop an optimization technique, $65\times$ faster than gradient descent on the same problem. We validate the effectiveness of SuperMix through extensive evaluations and ablation studies on two tasks of object classification and knowledge distillation. On the classification task, SuperMix provides the same performance as the advanced augmentation methods, such as AutoAugment. On the distillation task, SuperMix sets a new state-of-the-art with a significantly simplified distillation method. Particularly, in six out of eight teacher-student setups from the same architectures, the students trained on the mixed data surpass their teachers with a notable margin.