A Novel Spatio-Temporal-Wise Network for Action Recognition

Abstract

Action recognition is a challenging task that requires understanding the temporal relationships between frames. However, capturing and processing spatio-temporal and motion features is computationally expensive, making it difficult to apply to practical situations. We propose a novel approach called the Spatio-Temporal-Wise (STW) network to address this problem. The STW network inserts STW blocks, consisting of a Spatio-Temporal Fusion Module and a Temporal-Wise Module, into an existing 2D CNN. This approach requires very little additional computational overhead but brings huge performance improvements in recognizing human actions. The proposed method is evaluated on several public datasets, including Something-Something v1 & v2, Kinetics-400, UCF101, and HMDB51. STW achieved comparable or better performance on these datasets compared to state-of-the-art methods. Notably, the STW network improves recognition accuracy by 26.6% and 34.6% on the Something-Something v1 & v2 datasets, respectively, with less than 2% additional computational overhead. The results demonstrate that the STW network can significantly improve performance in action recognition tasks while requiring only a small additional computational overhead, which represents a promising direction for developing more efficient and effective approaches to handling temporal reasoning in action recognition, which may have important applications in the future.