Rendering Humans behind Occlusions.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Rendering 3D humans from monocular videos is challenging, especially with occlusions.
• Existing methods fail in real-world scenarios due to obstructions blocking the camera view.

Purpose of the Study:

• To develop a neural rendering approach for reconstructing 3D humans from occluded, in-the-wild monocular videos.
• To address limitations of current methods in handling partial occlusions.

Main Methods:

• Introduced Wild2Avatar, a novel neural rendering technique.
• Proposed occlusion-aware scene parameterization to decouple scene elements.
• Designed objective functions to enforce human separation and model completeness.

Main Results:

• Successfully rendered 3D humans from videos with significant occlusions.
• Demonstrated effectiveness on 14 challenging in-the-wild video datasets.
• Achieved robust human model reconstruction despite occlusions.

Conclusions:

• Wild2Avatar advances neural rendering for occluded monocular videos.
• The method enables 3D human reconstruction in complex, real-world environments.
• Paves the way for more robust human-aware scene understanding.