Human reconstruction using 3D Gaussian Splatting: a brief survey.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Traditional 3D human avatar reconstruction methods face challenges with computational cost, inference speed, and visual fidelity.
• Immersive applications like virtual reality (VR) and digital content creation demand high-quality, animatable 3D human representations.

Purpose of the Study:

• To review recent advancements in reconstructing high-fidelity and animatable 3D human avatars using 3D Gaussian Splatting (3DGS).
• To organize and discuss 3DGS-based reconstruction methods based on input data type (single-view vs. multi-view).

Main Methods:

• Leveraging parametric human models for canonical representation.
• Deforming 3D Gaussians using learnable skinning and novel regularization techniques.
• Employing hybrid Gaussian-mesh representations and specialized modules for complex geometry, clothing, occlusions, and fine details.

Main Results:

• 3DGS enables rapid training and real-time rendering (up to 361 FPS) for 3D human avatars.
• Advances include improved motion generalization, handling of complex clothing, efficient compression, and occlusion management.
• Methods are categorized by single-view and multi-view reconstruction approaches.

Conclusions:

• 3DGS offers a significant improvement over traditional methods for 3D human avatar reconstruction.
• Future research directions include further enhancements in handling complex scenarios and optimizing performance.
• The reviewed techniques are crucial for advancing immersive technologies and digital content creation.