Large Kernel Convolutional Neural Networks for Action Recognition Based on RepLKNet
Runlin Wang · 2023
To improve the classification accuracy of human actions in videos, two algorithms based on RepLKNet, a large kernel neural network, are proposed. By combining RepLKNet with Temporal Segment Networks (TSN) and incorporating the Temporal Shift Module (TSM), the models can better extract spatiotemporal information from videos, leading to improved classification accuracy. Experimental results demonstrate that the proposed models achieve higher classification accuracy compared to previous convolutional neural network (CNN) models. Moreover, these models outperform transformer-based models such as TimeSformer in classification accuracy while requiring significantly fewer parameters.