Human pose estimation with spatial context relationships based on graph convolutional network

Based on the statistical verification that there is a monotonic relationship between the probability of correctly estimating human joints and their confidence values, this paper proposed a human pose estimation model with spatial context relationships based on graph convolution. The model firstly selects credible joints according to the confidence threshold, and then uses human joints as nodes and the spatial context relationships between the joints as edges to construct a human pose graph, and achieves the spatial inference of credible joints to incredible joints by graph convolution. The model can be embedded in any existing pose estimator to achieve end-to-end training. And the experimental results on the MPII dataset show that the model can realize the propagation of position information between joints, and the spatial context relationships between joints are beneficial to correct the location of incredible joints.