Two-Path Motion Excitation for Action Recognition
Chichen Lin, Qi Wang, Xiao Han · 2023
Capturing discriminative motion information is the key to effective Spatio-temporal modeling for video motion recognition. 3D CNN can capture multi-frame motion information, but the computational effort is too large. Traditional 2D CNN is computationally small but powerless to capture motion information. Therefore, to solve this dilemma, we design a module that can be embedded in a 2D CNN. This module consists of two parallel short-term motion excitation (SME) paths and a long-term motion excitation (LME) path. The short-term motion excitation path is to compute feature variations between adjacent frames and use it to excite motion-sensitive channels in the image. The long-term motion excitation path calculates the feature changes among multiple frames, selects representative frame sequences, and aggregates motion information on time series with 3D convolution kernel. Compared with previous methods, this module focuses on solving the problem of poor motion information capture in long time series, while taking into account motion information capture in short time series. Experimental results show that the method has achieved good performance on benchmark data sets.