3D Convolutional Neural Network Pruning Algorithm for Action Recognition Based on Average Activation of RoI

• 1、School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Neural network pruning algorithms can help large models infer efficiently on resource-constrained devices with fewer number of parameters and computational complexity. However, existing pruning algorithms are often designed based on 2D convolutional networks, and the iterative pruning methods they use are costly in terms of time when applied to 3D convolutional networks. To address these problems, this paper proposes a pruning algorithm based on the average activation value of RoI regions based on the characteristics of action recognition 3D convolutional networks. The algorithm greatly reduces the time consumed in pruning by one-time pruning, and achieves a higher pruning effect than other common channel pruning algorithms by using the average activation of RoI (region of interest) regions in the feature map as a criterion to measure the importance of the filter. In addition, this paper also proposes a hybrid pruning strategy for the SlowFast algorithm, which has excellent performance in action recognition tasks, and obtains better pruning results by using different pruning algorithms for different branches.

Keywords： Neural Network Pruning Action Recognition 3D Convolutional Network