Speaker
Description
Efficient visualization of volumetric data produced by large-scale simulations requires compression to limit memory usage. We propose a lossy, multi-level compression technique designed for direct volume rendering of large scalar fields. Our approach represents the input volume as a hierarchy of grids containing polynomial coefficients, which adaptively approximate the fine-scale features found in the original voxel data. To enable efficient parallel construction, we employ Morton-order processing together with OpenVDB, and convert the resulting structure to NanoVDB for efficient sampling. We evaluate the proposed data structure by integrating it into Cycles renderer and testing it on several representative CFD and astrophysical datasets. The results demonstrate compression rates of 7.1× to 173× relative to uncompressed NanoVDB, with compression taking approximately 7 seconds per billion voxels. These results indicate that multi-level polynomial compression provides a practical balance between reducing memory consumption, preserving visual fidelity, and maintaining random access performance for large volumetric datasets in a Monte Carlo renderer.