Fixed-Budget Gaussian Encoding for Efficient Scalar Volume Compression

A novel technique has been developed for encoding scalar volumes as anisotropic Gaussian primitives within a predetermined budget. This method systematically allocates the entire set of primitives based on local field structures, which include aspects like position, orientation, and shape. It then refines directly against the scalar field without any need for densification, pruning, or alterations in count. The budget set beforehand dictates the storage for encoding, while the iteration schedule allows for a manageable refinement-time budget. In benchmark tests, truncation-aware field evaluation has shown to decrease encoding time by as much as 51 times, enabling 1.4 million Gaussians to encode a billion-voxel volume in under four minutes on a single desktop GPU, with reduced-iteration refinement finishing in under one minute. This technique was evaluated using five datasets totaling 2.1 million voxels. It effectively tackles the issue of storing, transferring, and loading scalar volumes generated more quickly than they can be processed, especially in scientific simulations where in situ reduction must operate within limited simulation resources. The research can be found on arXiv with the identifier 2608.14112.