Fundamental matrix estimation using generalized least squares

Classical approaches for estimating the fundamental matrix assume that Gaussian noise is contained in the estimates in view of mathematical tractability. However, this assumption will not be justified when the distribution computed is not normally distributed. We propose a new approach that does not make the Gaussian assumption, and so can attain robustness and accuracy in different conditions. The proposed framework, generalized least squares (GLS), is the extension of linear mixed-effects models considering the correlation between different data subsamples. We test the new model by using synthetic and real images, comparing it to the least median of squares technique.