Vision Transformers for Action Recognition: A Survey

Vision transformers are emerging as a powerful tool to solve computer vision problems. Recent techniques have also proven the efficacy of transformers beyond the image domain to solve numerous video-related tasks. Among those, human action recognition is receiving special attention from the research community due to its widespread applications. This article provides the first comprehensive survey of vision transformer techniques for action recognition. We analyze and summarize the existing and emerging literature in this direction while highlighting the popular trends in adapting transformers for action recognition. Due to their specialized application, we collectively refer to these methods as ``action transformers''. Our literature review provides suitable taxonomies for action transformers based on their architecture, modality, and intended objective. Within the context of action transformers, we explore the techniques to encode spatio-temporal data, dimensionality reduction, frame patch and spatio-temporal cube construction, and various representation methods. We also investigate the optimization of spatio-temporal attention in transformer layers to handle longer sequences, typically by reducing the number of tokens in a single attention operation. Moreover, we also investigate different network learning strategies, such as self-supervised and zero-shot learning, along with their associated losses for transformer-based action recognition. This survey also summarizes the progress towards gaining grounds on evaluation metric scores on important benchmarks with action transformers. Finally, it provides a discussion on the challenges, outlook, and future avenues for this research direction.