Presentation

· Contributors · Organizations · Search Program

NeuralVDB: High-resolution Sparse Volume Representation using Hierarchical Neural Networks

DescriptionNeuralVDB enhances the VDB framework for efficient sparse volumetric data storage by integrating machine learning. This new structure significantly reduces memory usage while maintaining flexibility with minimal compression errors. It combines a shallow VDB tree with hierarchical neural networks for topology and value encoding, achieving high compression ratios and outperforming other neural representations. Additionally, NeuralVDB improves animation compression and temporal coherence through warm-starting from previous frames.

Authors

Doyub Kim

NVIDIA, USA

Minjae Lee

NVIDIA, USA