Structure recovery from uncalibrated images using iterative perspective factorization

Structure recovery is a widely used data acquisition way for scene analysis. Factorization based methods have become the mainstream methods for structure recovery these years, which generally recover structures by solving the rectification matrix. However, these solutions will be invalid when eigen-decomposing a negative definite rectification matrix. Hence, we present an improved iterative perspective factorization method in which structure and motions are directly solved by iteratively imposing constraints rather than matrix decomposition. Experiments on both synthetic data and real images show that the proposed method can avoid the invalidation of general factorization-based solutions caused by failures and efficiently recover the geometric structure, camera motion parameters and camera focal lengths simultaneously from uncalibrated images with fast convergence rate, high accuracy and noise tolerance.