Video Transformers: A Survey

Transformer models have shown great success modeling long-range interactions. Nevertheless, they scale quadratically with input length and lack inductive biases. These limitations can be further exacerbated when dealing with the high dimensionality of video. Proper modeling of video, which can span from seconds to hours, requires handling long-range interactions. This makes Transformers a promising tool for solving video related tasks, but some adaptations are required. While there are previous works that study the advances of Transformers for vision tasks, there is none that focus on in-depth analysis of video-specific designs. In this survey we analyse and summarize the main contributions and trends for adapting Transformers to model video data. Specifically, we delve into how videos are embedded and tokenized, finding a very widspread use of large CNN backbones to reduce dimensionality and a predominance of patches and frames as tokens. Furthermore, we study how the Transformer layer has been tweaked to handle longer sequences, generally by reducing the number of tokens in single attention operation. Also, we analyse the self-supervised losses used to train Video Transformers, which to date are mostly constrained to contrastive approaches. Finally, we explore how other modalities are integrated with video and conduct a performance comparison on the most common benchmark for Video Transformers (i.e., action classification), finding them to outperform 3D CNN counterparts with equivalent FLOPs and no significant parameter increase.