GPMN: Human Pose Explicit Modeling Network based on Graph Model

The human pose estimation task requires the use of visual cues and anatomical relationships between joints to locate key points. Due to the structural dependencies between human joints, it is difficult to model the dependencies between joints with a convolutional neural network-based approach. In this paper, we propose a graph pose modeling network (GPMN) based on a graph model, which constructs a structural graph model between joints and then uses a convolutional neural network to explicitly model the dependencies between joints. Specifically, this method extracts features through a backbone network and uses a graph model to obtain proximity joint dependencies. At the same time, a cross-cross-attention mechanism is used to obtain the distant joint dependencies for reinforcing challenging key points. To solve the problem that the challenging key points of the network are weakened during the training process, this paper uses a focal loss function to make the network focus more on the challenging key points, such as knees and ankles. Under the same experimental conditions, HRNet, the highest accuracy feature extraction network, and ResNet, the classical feature extraction network, are used as the backbone network. The accuracy of the algorithm with HRNet network as the backbone is improved by 2.6% compared with the original network.