Advances in Few-Shot Action Recognition: A Comprehensive Review

Action recognition holds a key position in the field of video understanding. While methods based on deep learning have made significant strides in big data backgrounds, their application in few-shot scenarios for action recognition still faces numerous challenges. This review delves deeply into over 20 methods in the few-shot action recognition (FSAR) area from recent years. Initially, following the research's developmental trajectory, we systematically categorize the current mainstream methods for few-shot action recognition into three perspectives: memory networks, metric learning, and data augmentation. In order to ensure a fair comparison, we examine the performance of several approaches on multiple datasets. Lastly, we offer insights into the future development directions of this field. This review is the first review in the FSAR domain, summarizing a wealth of research outcomes, techniques, and theories. It offers researchers a cohesive perspective and illuminates potential new avenues for innovative studies.