Learning Representational Invariances for Data-Efficient Action Recognition

Data augmentation is a ubiquitous technique for improving image classification when labeled data scarce. Constraining the model predictions to be invariant diverse augmentations effectively injects desired representational invariances (e.g., invariance photometric variations) and helps improve accuracy. Compared data, appearance variations in videos are far more complex due additional temporal dimension. Yet, methods remain under-explored. This paper investigates various strategies that capture different video invariances, including photometric, geometric, temporal, actor/scene augmentations. When integrated with existing semi-supervised learning frameworks, we show our strategy leads promising performance on Kinetics-100/400, Mini-Something-v2, UCF-101, HMDB-51 datasets low-label regime. We also validate fully supervised setting demonstrate improved performance.