A Novel Spatio-Temporal-Wise Network for Action Recognition

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