Modeling Internal Object Structures with 3D Gaussian Splatting

Develop a method within the 3D Gaussian Splatting (GS) framework that accurately models the internal volumetric structure of objects, such as those in computed tomography, by organizing Gaussian primitives to align with true object interiors and enabling volumetric representations suitable for downstream applications.

Background

The survey highlights that while 3D Gaussian Splatting excels at realistic surface rendering, its current splatting and density control processes produce an unorganized representation that does not align well with true internal object structures. Many applications, particularly in medical imaging such as computed tomography, require volumetric modeling rather than surface-only representations.

Existing attempts partially address related needs: Li et al. (2023) use 3D Gaussians with density control as a basis for volumetric representation without involving splatting, while X-Gaussian (2024) retains splatting for fast training and inference but cannot generate volumetric representations. This leaves a gap for a GS-based approach that both supports volumetric outputs and maintains the practical performance benefits of splatting.