Markerless 3D reconstruction of temporally deforming surfaces from multiple cameras

3D content generation is a major challenge in computer graphics, and generating realistic objects with time-varying shape and motion is a big part of creating realistic virtual environments. An attractive approach for generating content is to acquire the shape of real objects using computer vision techniques. Capturing the real world avoids the complexity of mathematical simulation as well as the time-consuming and difficult process of creating content manually. Traditional acquisition methods, such as range scanning, work well for static objects. However, temporally deformable objects like garments and human faces pose greater challenges. Here, the capture process must reconstruct both the time-varying shape as well as the motion of the object. Additionally, per-frame texture maps should be acquired for realistic rendering. Previous methods contain various limitations, such as the use of structured illumination patterns, or hand-placed markers on the surface in order to guide the reconstruction. Placing markers on garments limits the types of garments that can be captured. Both markers and structured light on human faces limit the ability to capture high-quality per-frame texture maps, which are essential for reconstructing performance-specific details such as sweating, wrinkles and blood flow. In this thesis we propose new, entirely passive techniques for advancing the state-of-the-art in 3D reconstruction of temporally deformable surfaces. We reconstruct high-resolution, temporally compatible geometry using multiple video cameras, without the need for either markers or structured light patterns. This thesis contains several contributions, both in computer graphics as well as computer vision. We start with new methods for multi-camera calibration and synchronization, which allow us to employ inexpensive consumer video cameras for reconstruction. We also present a novel technique for multi-view stereo reconstruction, capable of building high-resolution geometric meshes from a number of images. Finally, we combine these ideas and provide two inexpensive solutions for high-resolution 3D reconstruction of temporally deforming surfaces without the need for markers. The first is an application for capturing the shape and motion of garments, and the second is a new method for acquiring facial performances of actors. These new techniques have the potential for high impact in the film and video game industries.Science, Faculty ofComputer Science, Department ofGraduat