Hybrid Attention MLP-Graph Network for 3D Human Pose Estimation
Feiyue Qiu, Lin Sun, Delong Peng, Jian Qiang Zhou · 2024
To solve the problems of complex background, self-occlusion, and posture inaccuracy detection in 3D human pose estimation, a hybrid attention multilayer perceptron graph convolutional network (HAMLP-Graph) was proposed. Based on an improved multi-layer perceptron mixer (MLP-Mixer) architecture, the network introduces spatial and channel attention mechanisms to enhance useful spatial channel information by adaptively adjusting the weights of different locations and channels, enabling the network to identify key human feature points. In addition, the graph convolution module is added to the network and the connection and symmetry relations of human bone nodes are fully utilized. The features of nodes and edges are extracted by using convolution operations on the graph structure, to improve the accuracy of the model for node detection. The experimental results show that compared with GraphMLP, the MPJPE index on the Human3.6M dataset decreased by 3.5 percentage points. On the MPI-INF-3DHP dataset, 3DPCK index and AUC index increased by 1.7 percentage points and 2.1 percentage points respectively, which verified that the proposed network was more robust.