Video Swin Transformer

TLDR

The vision community is shifting from CNNs to Transformers, with pure Transformer architectures achieving top accuracy on major video recognition benchmarks, yet these models globally connect patches across spatial and temporal dimensions. The study proposes adding locality to video Transformers to improve the speed‑accuracy trade‑off over globally‑attentive models. The authors adapt the image‑domain Swin Transformer, preserving locality and leveraging pre‑trained image models. The approach attains state‑of‑the‑art accuracy on multiple benchmarks, achieving 84.9 % top‑1 on Kinetics‑400, 85.9 % on Kinetics‑600 with ~20× less pre‑training data and ~3× smaller model size, and 69.6 % on Something‑Something v2.

Abstract

The vision community is witnessing a modeling shift from CNNs to Transformers, where pure Transformer architectures have attained top accuracy on the major video recognition benchmarks. These video models are all built on Transformer layers that globally connect patches across the spatial and temporal dimensions. In this paper, we instead advocate an inductive bias of locality in video Transformers, which leads to a better speed-accuracy trade-off compared to previous approaches which compute self-attention globally even with spatial-temporal factorization. The locality of the proposed video architecture is realized by adapting the Swin Transformer designed for the image domain, while continuing to leverage the power of pre-trained image models. Our approach achieves state-of-the-art accuracy on a broad range of video recognition benchmarks, including on action recognition (84.9 top-l accuracy on Kinetics-400 and 85.9 top-l accuracy on Kinetics-600 with ~20× less pre-training data and ~3× smaller model size) and temporal modeling (69.6 top-l accuracy on Something-Something v2).