iVR-GS: Inverse Volume Rendering for Explorable Visualization via Editable 3D Gaussian Splatting.

Area of Science:

• Computer Graphics
• Scientific Visualization
• Artificial Intelligence

Background:

• Volume visualization enables interactive 3D data exploration using transfer functions (TF) and lighting.
• Real-time rendering of large volumes requires significant GPU power and memory.
• Existing novel view synthesis (NVS) methods offer faster rendering but limit user exploration due to fixed TF settings.

Purpose of the Study:

• To introduce inverse volume rendering via Gaussian splatting (iVR-GS) for efficient and interactive volume exploration.
• To enable scene editing capabilities within NVS methods for enhanced data interpretation.
• To reduce rendering costs and hardware requirements for large-scale volume visualization.

Main Methods:

• Developed iVR-GS, an NVS method utilizing Gaussian splatting for volume rendering.
• Composed multiple iVR-GS models, each with basic TFs for disjoint visible parts, to render the entire volumetric scene.
• Each iVR-GS model comprises 3D editable Gaussians for real-time rendering and scene manipulation.

Main Results:

• Demonstrated superior reconstruction quality of iVR-GS compared to Plenoxels, CCNeRF, and base 3DGS.
• Showcased the composability of iVR-GS for rendering complex volumetric scenes.
• Validated iVR-GS performance on various volume datasets, highlighting its efficiency and editability.

Conclusions:

• iVR-GS offers a powerful solution for interactive volume exploration with reduced rendering costs.
• The method enables real-time scene editing, significantly enhancing user interaction with 3D data.
• iVR-GS represents a significant advancement over existing NVS techniques for volume visualization.