[2B1-OS-41d] OS-41

[2B1-OS-41d-01] MPM-Based GNN Learning Simulator for 3DGS Animation

〇Tomohiro Mimura¹, Yoko Inaba², Shohei Taniguchi³, Yutaka Matsuo³ (1. Graduate School of Engineering, Kyoto University, 2. Faculty of Engineering, Kyoto University, 3. Graduate School of Engineering, The University of Tokyo)

Keywords:3DGS, Material Point Method, Physics, Animation, Graph Neural Network

This paper explores whether a GNN-based Graph Network Simulator can accelerate the Material Point Method within a 3D Gaussian Splatting pipeline for 4D animation generation. We investigate GNS's ability to efficiently predict MPM particle positions and deformation gradients, aiming to speed up the simulation process. Our contributions include: extending GNS for simultaneous tensor prediction in 3DGS, identifying effective feature combinations for predicting positions and deformation gradients, and evaluating the viability of GNS for physics simulation in 3DGS. While using deformation gradients as input features did reduce training losses, incorporating local velocity gradients or temporal differences of the deformation gradients increased prediction errors. Furthermore, GNS was slower than GPU-accelerated MPM in realistic 3DGS simulations. This suggests that GNS in its current configuration is not advantageous for single-step predictions in 3DGS.