Multiview Transformers for Video Recognition

TLDR

Video understanding requires reasoning at multiple spatiotemporal resolutions, yet transformer architectures have not explicitly modeled these varying resolutions. The authors introduce Multiview Transformers for Video Recognition (MTV) to address this gap. MTV employs separate encoders for different video views with lateral connections that fuse information across views, enabling multi‑resolution reasoning. Ablation studies show MTV outperforms single‑view counterparts in accuracy and efficiency, and it achieves state‑of‑the‑art performance on six datasets, further improved by large‑scale pretraining. Code and checkpoints are available at https://github.com/google-research/scenic.

Abstract

Video understanding requires reasoning at multiple spatiotemporal resolutions – from short fine-grained motions to events taking place over longer durations. Although transformer architectures have recently advanced the state-of-the-art, they have not explicitly modelled different spatiotemporal resolutions. To this end, we present Multiview Transformers for Video Recognition (MTV). Our model consists of separate encoders to represent different views of the input video with lateral connections to fuse information across views. We present thorough ablation studies of our model and show that MTV consistently performs better than single-view counterparts in terms of accuracy and computational cost across a range of model sizes. Furthermore, we achieve state-of-the-art results on six standard datasets, and improve even further with large-scale pretraining. Code and checkpoints are available at: https://github.com/google-research/scenic.