Achieving view-distance and -angle invariance in motion prediction using a simple network.

Abstract

Recently, human motion prediction has gained significant attention and achieved notable success. However, current methods primarily rely on training and testing with ideal datasets, overlooking the impact of variations in the viewing distance and viewing angle, which are commonly encountered in practical scenarios. In this study, we address the issue of model invariance by ensuring robust performance despite variations in view distances and angles. To achieve this, we employed Riemannian geometry methods to constrain the learning process of neural networks, enabling the prediction of invariances using a simple network. Furthermore, this enhances the application of motion prediction in various scenarios. Our framework uses Riemannian geometry to encode motion into a novel motion space to achieve prediction with an invariant viewing distance and angle using a simple network. Specifically, the specified path transport square-root velocity function is proposed to aid in removing the view-angle equivalence class and encode motion sequences into a flattened space. Motion coding by the geometry method linearizes the optimization problem in a non-flattened space and effectively extracts motion information, allowing the proposed method to achieve competitive performance using a simple network. Experimental results on Human 3.6M and CMU MoCap demonstrate that the proposed framework has competitive performance and invariance to the viewing distance and viewing angle.