AZY-GCN: Multi-scale feature suppression attentional diagram convolutional network for human pose prediction

Due to the randomness and non-periodic nature of the future posture of the human body, the prediction of the posture of the human body has always been a very challenging task. In the latest research, graph convolution is proved to be an effective method to capture the dynamic relationship between the human body posture joints, which is helpful for the human body posture prediction. Moreover, graph convolution can abstract the pose of the human body to obtain a multi-scale pose set. As the level of abstraction increases, the posture movement will become more stable. Although the average prediction accuracy has improved significantly in recent years, there is still much room for exploration in the application of graph convolution in pose prediction. In this work, we propose a new multi-scale feature suppression attention map convolutional network (AZY-GCN) for end-to-end human pose prediction tasks. We use GCN to extract features from the fine-grained scale to the coarse-grained scale and then from the coarse-grained scale to the fine-grained scale. Then we combine and decode the extracted features at each scale to obtain the residual between the input and the target pose. We also performed intermediate supervision on all predicted poses so that the network can learn more representative features. In addition, we also propose a new feature suppression attention module (FISA-block), which can effectively extract relevant information from neighboring nodes while suppressing poor GCN learning noise. Our proposed method was evaluated on the public data sets of Human3.6M and CMU Mocap. After a large number of experiments, it is shown that our method has achieved relatively advanced performance.