Super 3D from Video

We propose to develop a novel approach to extracting 3D structure from wide-area video with sub-pixel accuracy. Multiple redundant video frames from very similar viewpoints enable super-resolution, while multiple frames from separated viewpoints enable 3D structure estimation. In our approach we will combine elements of the two processes through a variant of space carving, while avoiding the challenge of sub-pixel registration required for super-resolution. Our approach is based on voxel removal, where 3D volumetric elements are removed or “carved” out of an initial cube when their appearance is not consistent with multiple images. We will enhance space carving with edge-based analysis. Edges provide sub-pixel localization in each image, and by combining highly localized edges on occluding contours from different viewpoints we can achieve sub-pixel localization in 3D. Our approach can model and leverage the characteristics of the sensor, such as rolling shutters.