Abstract:
Self-attention learns pairwise interactions to model long-range dependencies, yielding great improvements for video action recognition. In this paper, we seek a deeper understanding of self-attention for temporal modeling in videos. We first demonstrate that the entangled modeling of spatio-temporal information by flattening all pixels is sub-optimal, failing to capture temporal relationships among frames explicitly. To this end, we introduce Global Temporal Attention (GTA), which performs global temporal attention on top of spatial attention in a decoupled manner. We apply GTA on both pixels and semantically similar regions to capture temporal relationships at different levels of spatial granularity. Unlike conventional self-attention that computes an instance-specific attention matrix, GTA directly learns a global attention matrix that is intended to encode temporal structures that generalize across different samples. We further augment GTA with a cross-channel multi-head fashion to exploit channel interactions for better temporal modeling. Extensive experiments on 2D and 3D networks demonstrate that our approach consistently enhances temporal modeling and provides state-of-the-art performance on three video action recognition datasets.