Lightweight skeletal key point extraction algorithm based on TransPose network
Wenxian Zeng, Weiguang Li, Yuesong Li · 2023
TransPose achieves excellent results in human pose estimation tasks, but it has disadvantages such as a lot of network parameters, high computational complexity of the model, and slow convergence speed. The TransPose model is improved to address the above problems. First, we use the Ghost module to change the ordinary convolutional structure in the convolutional backbone network HRNet to decrease the quantity of model parameters and calculation complexity; second, we use the multi-head mechanism of attention to increase the encoder learnable information and improve the model accuracy; third, we introduce the normalization structure of the front layer to accelerate the model convergence. It is proved by experimental results that the amount of computation of the improved TransPose model on COCO dataset is reduced from 21.8G to 15G, the number of parameters is reduced from 17.5M to 4.85M, and the model converges faster while maintaining a high accuracy rate.