Multiscale Models for Real-Time Human Pose Estimation
Zongyou Liu, Wenbai Li, Hongtao Wang · 2023
Although the lightweight OpenPose uses the depth-wise separable convolution to form the backbone to improve the computing efficiency of the model on the CPU, the lightweight OpenPose also has the following shortcomings. (1) The model has weak feature processing ability. (2) The model cannot identify multi-scale features well. In order to solve the above problems, a multi-scale model is proposed. We adjusted the network structure of the lightweight OpenPose model with reference to the network structure of RES2NET. The modified model not only recognizes multi-scale features but also has strong feature processing capabilities. The AP value of our model on the COCO2017 validation set is 0.412. The AP value of Multiscale Models is 0.037 higher than lightweight OpenPose on the COCO2017 validation set. Our future direction is to discover the influence of the model's ability to extract multi-scale features on the model AP value.