VRGaussianAvatar: Integrating 3D Gaussian Avatars into VR.

Area of Science:

• Computer Graphics
• Virtual Reality
• Human-Computer Interaction

Background:

• Real-time avatar generation is crucial for immersive virtual reality (VR).
• Existing methods often struggle with full-body representation and rendering efficiency.
• 3D Gaussian Splatting (3DGS) offers high-fidelity rendering but requires efficient integration into VR.

Purpose of the Study:

• To develop an integrated system for real-time, full-body 3D Gaussian Splatting (3DGS) avatars in VR.
• To enable avatar generation using only head-mounted display (HMD) tracking signals.
• To improve the efficiency and visual quality of stereoscopic rendering for VR avatars.

Main Methods:

• A parallel pipeline with a VR Frontend for pose estimation (inverse kinematics) and a GA Backend for 3DGS rendering.
• Stereoscopic rendering of 3DGS avatars from single-image reconstructions.
• Introduction of Binocular Batching for efficient joint processing of stereo views.

Main Results:

• VRGaussianAvatar sustains interactive VR performance.
• The system achieves higher perceived appearance similarity, embodiment, and plausibility compared to mesh avatars.
• Quantitative performance tests and user studies validate the system's effectiveness.

Conclusions:

• VRGaussianAvatar provides a novel solution for real-time, high-fidelity full-body avatars in VR.
• The proposed Binocular Batching significantly improves rendering efficiency for high-resolution VR displays.
• The system enhances the immersive experience by improving avatar realism and user embodiment.