Human Pose Estimation: Multi-stage Network Based on HRNet
Xiaodong Ji, Qiaoning Yang, Xiuhui Yang, Jiahao Zheng, Mengyan Gong · Journal of Physics Conference Series · 2022
Abstract Multi-stage network uses stacked networks to enhance the feature extraction capability, and can gradually refine the keypoints with the information of previous stages’ output. Obviously, multi-stage networks are more suitable for human pose estimation. However, most current multi-stage networks use a codec structure as the backbone in which downsample will cause information loss. HRNet maintains high-resolution features to supply the information which is lost in down-sampling stage. In this regard, we propose a novel two-stage network with HRNet as the backbone and stacked codec structure. HRNet has more efficient feature extraction capability, and the stacked codec network can utilize the multi-scale features generated by HRNet more effectively. This method obtains a 1.2AP improvement compared to HRNet and a significant improvement compared to other two-stage networks.