Abstract:

Structure-and-Motion is one of the central problems in Geometric Computer Vision. Its goal is to reconstruct 3D information from multiple input images, while estimate the relative camera motions across the images. When there exists only one relative motion between two images, the problem has been intensively studied and well solved. However, when there are multiple motions in presence, it becomes substantially hard, due to its well-known chicken-and-egg type character. In this talk I will describe a new Linear Programming-based method to solve this problem effectively. Also, I will demonstrate how this method offers a principled paradigm for solving other similar problems in computer vision and machine learning.