Multilevel Joint Association Networks for Diverse Human Motion Prediction

Predicting accurate and diverse human motion presents a challenging task due to the complexity and uncertainty of future human motion. Existing works have explored sampling techniques and body modeling approaches to enhance diversity while maintaining the accuracy of human motion prediction. However, most of them often fall short in capturing the hierarchical features of the correlations between joints. To address these limitations, we propose in this paper the Multilevel Joint Association Network, a novel deep generative model designed to achieve both diverse and controllable motion prediction by adjusting the manner in which the human body is modeled. Our model incorporates two Graph Convolution Networks (GCNs) to enhance the extraction of features, resulting in more accurate future motion samples. Furthermore, we employ a multi-level Transformer generator that effectively capture the contact information between human joints and the personality characteristics of human joints, enabling the generated future motion samples with high diversity and low error. Extensive experimental results on two challenging datasets Human3.6 M and HumanEva-I, indicate that the proposed method achieves state-of-the-art performance in terms of both diversity and accuracy.