Head Avatar Reconstruction from Large-volume Video Based on Mesh-embedded Gaussians

Accurate simulation of facial dynamics, expressions, and viewpoint changes is a key challenge when building realistic 3D digital humans, especially when reconstructing details such as teeth, mouths, and eyes. To efficiently reconstruct high-fidelity head avatar, we propose MEGAvatar, which takes large-volumn facial video as input and combines explicit mesh and implicit Gaussians as 3D representation to reconstruct dynamic head. This approach separates the deformation and appearance of the digital head, enabling accurate restoration of complex facial details (e.g., teeth, eyes, and lips). We also introduce an innovative single-frame 3D Gaussian extraction strategy to reduce the influence of incorrect poses on the model and ensure controllable Gaussian motion. A head video of a person is about 500MB, with about 3,000 frames. To build a virtual conference room containing several digital heads, we will use an edge computing system. The edge node is responsible for reconstructing the digital human, converting the original video data into a compact Gaussian representation (about 100 MB), and then transmitting it to the central node. At the central node, the system integrates multiple virtual characters into a large virtual environment while ensuring consistency and scale matching between participants. Compared with existing methods, our method reconstructs high-fidelity digital heads, promoting the widespread application of immersive high-fidelity digital humans in virtual reality.