Group RandAugment: Video Augmentation for Action Recognition

Data augmentation, as a critical strategy in deep learning, well improves the sample diversity for network training, leading to the obvious improvement of model generalization ability. Besides, automatic data augmentation, while sparking in image tasks, has attracted too little attention in video recognition task. Therefore, this work explores a novel group random augmentation (GRA) that automatically augments video data for recognition. GRA first collects 21 augmentation transformations to enrich the sample diversity, and then divides these transformations into four groups (i.e., pixel-transform, rigidtransform, erasing-transform, environment-transform) to reduce the excessive regularization caused by similar augmentation. To reduce the computational cost of finding the optimal setting, GRA adopts the combination form to select the optimal situation. For frames in the same video clip, GRA uses synchronous GRA. Additionally, the proposed GRA can be integrated into any existing video frameworks. To prove the effectiveness of GRA, we conduct experiments on two commonly used video action recognition benchmarks (Something-Something V1 & V2) and three typical frameworks (TSM, GSM, and TEA), whose results demonstrate GRA can improve performance in all cases without adding additional computational cost.