Video Behavior Recognition Model Based on 3D Reprameterized Convolution
Ping Liu, Sheng Quan Xie, Chaofu Lin, Hao Wu · 2023
In addressing the challenge posed by the extensive parameters of 3D convolutional modules, this paper extends the 2D reparameterization method to propose a 3D reparameterized convolutional module. During training, this module capitalizes on a multi-path structure for effective feature learning, while simplifying into a single-path structure during deployment to curtail computational costs. Moreover, to enhance the model's focus on the temporal dimension, a channel-time attention module is introduced, further augmenting the model's accuracy. Leveraging these modules, a reparameterized 3D convolutional neural network is developed for video behavior recognition. Experimental results underscore the efficacy of the proposed modules, showcasing improved model accuracy and reduced inference time.