ConvCCpose: Learning Coordinate Classification Tokens for Human Pose Estimation Based on ConvNeXt
Zhi Liu, Cong Chen, Hailan Xu, Shengzhao Hao, Lei Liu · 2023
Human pose estimation plays a vital role in real-world applications. In academia, 2D HPE has achieved excellent performance, but its application in the industry is challenging due to issues such as a large number of model parameters and high latency. To address this, this paper presents an efficient human pose estimation model called ConvCCPose, based on a convolutional neural network architecture. ConvCCPose tackles the challenge by separating the coordinate localization task into a classification task for horizontal and vertical coordinates, eliminating the need for additional post-processing and facilitating easier deployment. The ConvNeXt serves as the backbone for feature extraction, while a spatial bias module is introduced to enhance the backbone's feature extraction capability. Additionally, a contextual feature aggregation module is designed to enable multi-scale feature fusion. This module aids in extracting human structural representations that are valuable for accurate pose estimation. Specifically, on the MSCOCO validation dataset, ConvCCPose obtains 73.8% AP with 2.8 GFLOPs. It achieves a balance between computational efficiency and accuracy, making it more suitable for practical applications.