Human pose estimation: An improvement based on high resolution
Wanjun Yu, Huiyu Chen · 2022
Aiming at the large number of parameters in the high-resolution network and the complex operation and easy to cause information loss in the acquisition of low-resolution feature maps, this paper proposes an improved high-resolution human key point detection network DCANet(Dense Coordinate Attention Net). With HRNet as the backbone network, DenseNet was used for lightweight improvement, so as to reduce the operation parameters of the network and optimize the operation speed of the network. At the same time, CA(Coordinate Attention) was added after obtaining low-resolution branches ) module, which extracts channel and location information to overcome the information loss caused by repeated down-sampling operations in the process of acquiring low-resolution feature maps, It can reduce network complexity and ensure accuracy at the same time.