From Meshes to Splats: Exploring Dynamic Gaussian Splatting for Human Avatars in Volumetric Capture Workflows

Volumetric capture has revolutionized immersive media by reconstructing virtual scenes from real-world footage, captured through multi-camera setups. This method preserves authentic movements and textures without relying on computer-generated elements. While mesh-based representations have dominated volumetric pipelines, their limitations in handling complex scenarios restrict creative freedom and impose strict capture guidelines. Gaussian Splatting introduces a novel approach for real-time radiance field rendering, using 3D Gaussians to represent scenes. The core purpose of this thesis is to investigate Dynamic Gaussian Splatting as a potential alternative to meshes in Volucap GmbH’s volumetric capture workflow, focusing on human avatars rendering for extended reality (XR) experiences. This research explores two primary implementation approaches for Dynamic Gaussian Splatting: frame-by-frame Gaussian Splatting and 4D Gaussian Splatting. The advantages and challenges of both these methods are analyzed to determine the most suitable approach for integration into the company’s existing workflow. This study then delves into a comparison between mesh-based and Gaussian-based outputs both across standard scenarios, where meshes generally perform well, and complex cases that challenge traditional pipelines, such as transparencies, thin fabrics and objects, dark clothing, uniform textures, reflective surfaces and hair reconstruction. The evaluation integrates quantitative and qualitative analysis, assessing Gaussian Splatting’s ability to overcome mesh limitations while maintaining Volucap’s production standards. The findings will help determine the suitability of this technique for a potential integration into Volucap’s workflow.