CVPR2025

Nonisotropic Gaussian Diffusion for Realistic 3D Human Motion Prediction

Cecilia Curreli, Dominik Muhle, Abhishek Saroha, Zhenzhang Ye, Riccardo Marin, Daniel Cremers

摘要

P as t G T F u tu re P re d ic te d fu tu re s C lo se st D iv er se #1 D iv er se #2 D iv er se #3 Figure 1. SkeletonDiffusion generates futures that are simultaneously diverse and realistic. With a nonisotropic diffusion formulation reflecting the skeleton structure, we predict motions that are plausible and semantically coherent with the input past while being highly diverse. Here, we show the most diverse ensemble of three motions including the prediction closest to the ground truth among 50 generated futures.