NeuMA: Neural Material Adaptor for Visual Grounding of Intrinsic Dynamics

Authors: Junyi Cao, Shanyan Guan, Yanhao Ge, Wei Li, Xiaokang Yang, Chao Ma

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments on various dynamics in terms of grounded particle accuracy, dynamic rendering quality, and generalization ability demonstrate that Neu MA can accurately capture intrinsic dynamics.
Researcher Affiliation Collaboration 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 vivo Mobile Communication Co., Ltd. {junyicao, xkyang, chaoma}@sjtu.edu.cn {guanshanyan, halege, liwei.yxgh}@vivo.com
Pseudocode Yes Algorithm 1: Particle Binding
Open Source Code Yes Project Page: https://xjay18.github.io/projects/neuma.html.
Open Datasets Yes For a comprehensive evaluation, we consider both synthetic and real-world data. Synthetic Data. Following PAC-Ne RF [45] and NCLaw [53], we use MPM [34] to simulate 6 kinds of dynamics... The generated 6 benchmarks are named Bouncy Ball , Jelly Duck , Rubber Pawn 1, Clay Cat , Honey Bottle , and Sand Fish . We report the simulation details in Appendix B. Real-world Data. We adopt the real-world data provided by Spring-Gaus [96] to assess the visual grounding performance in the real world.
Dataset Splits Yes To ensure an accurate object appearance and geometry, we additionally render 50 uniformly sampled viewpoints at t = 0, with the cameras evenly spaced on a sphere covering the object of interest, and use these data for static reconstruction.
Hardware Specification Yes All the experiments are conducted on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions software like Blender [9], RAdam optimizer [49], and MPM [34], but it does not provide specific version numbers for these software components or libraries.
Experiment Setup Yes We optimize the neural material adaptor Mθ using RAdam [49] optimizer with a cosine learning rate scheduler for 1, 000 iterations for each scene. We choose τ = 95% and ρ = 1, 000 for all experiments unless otherwise specified. Our MPM simulator operates in a [0, 1]3 cube with a fixed resolution of 323 and 703 for synthetic and real-world data unless otherwise specified.