From Public Wiki
This project concerned "understanding" 2-D building facades in terms of features such as windows, doors, and other architectural elements. We studied techniques for constructing a "clean" texture map of a partially occluded building facade both from a series of images taken from a moving camera and a single still image. In the first case, building regions blocked by trees, signs, people, and other foreground objects are recovered using parallax and robust filtering techniques. In the case of a single image, the key assumption that the background is strongly structured allows automatic detection of occluders as outlier pixels in near-regular textures and their replacement via a robust subspace reconstruction process driven by tile appearance statistics.
This work was described most comprehensively in Thommen Korah's PhD thesis from 2007. Aspects of the method were reported in ECCV 2008, IEEE Trans. Image Processing 2007, ICIP 2007, ISVC 2006, ACCV 2006, and ICIP 2005.
