Robust Multiple-View Geometry Estimation Based on GMM.

MX Hu,Q Xing,BZ Yuan,XF Tang

2002-01-01

Abstract:Given three partially overlapping views of the scene from which a set of point or line correspondences have been extracted, 3D structure and camera motion parameters can be represented by the trifocal tensor, which is the key to many problems of computer vision on three views. Unlike in conventional typical methods, the residual value is the only rule to eliminate outliers with large value, we build a Gaussian mixture model assuming that the residuals corresponding to the inliers come from Gaussian distributions different from that of the residuals of outliers. Then Bayesian rule of minimal risk is employed to classify all the correspondences using the parameters computed from GMM. Experiments with both synthetic data and real images show that our method is more robust and precise than other typical methods because it can efficiently detect and delete the bad corresponding points, which include both bad locations and false matches.