Spatial-and-Temporal-Weighted Structure from Motion

The paper focuses on the problem of structure from motion and proposes a spatial-and-temporal-weighted factorization algorithm. The contributions of the paper are as follows: First, it is demonstrated that the image reprojection error is generally in proportion to the error magnitude contained in the image. Second, the error distribution contained in the measurement matrix is modelled spatially and temporally from the image reprojection residuals. Third, based on the estimated error distribution, a spatial-and-temporal-weighted factorization algorithm is proposed to handle the uncertainty caused by significant image noises. Compared with existing approaches, the proposed technique does not require a prior information of image measurement, and it is easy to implement. Extensive experiments on synthetic data and real images are performed to validate the proposed method.