Inverse rendering for scene reconstruction in general environments

Abstract

Demand for high-quality 3D content has been exploding recently, owing to the advances in 3D displays and 3D printing. However, due to insufficient 3D content, the potential of 3D display and printing technology has not been realized to its full extent. Techniques for capturing the real world, which are able to generate 3D models from captured images or videos, are a hot research topic in computer graphics and computer vision. Despite significant progress, many methods are still highly constrained and require lots of prerequisites to succeed. Marker-less performance capture is one such dynamic scene reconstruction technique that is still confined to studio environments. The requirements involved, such as the need for a multi-view camera setup, specially engineered lighting or green-screen backgrounds, prevent these methods from being widely used by the film industry or even by ordinary consumers. In the area of scene reconstruction from images or videos, this thesis proposes new techniques that succeed in general environments, even using as few as two cameras. Contributions are made in terms of reducing the constraints of marker-less performance capture on lighting, background and the required number of cameras. The primary theoretical contribution lies in the investigation of light transport mechanisms for high-quality 3D reconstruction in general environments. Several steps are taken to approach the goal of scene reconstruction in general environments. At first, the concept of employing inverse rendering for scene reconstruction is demonstrated on static scenes, where a high-quality multi-view 3D reconstruction method under general unknown illumination is developed. Then, this concept is extended to dynamic scene reconstruction from multi-view video, where detailed 3D models of dynamic scenes can be captured under general and even varying lighting, and in front of a general scene background without a green screen. Finally, efforts are made to reduce the number of cameras employed. New performance capture methods using as few as two cameras are proposed to capture high-quality 3D geometry in general environments, even outdoors.