From: Harry L. Voorhees <hlv@st...>  20060714 18:03:21

Jun Liu wrote: > Hi, > > I am using "FMatrixComputeRANSAC" in MVL (part of OXL) to compute fundamental > matrix based on detected feature points correspondences. There is very weird > output though: given exactly the same input data, it will produce different > fundamental matrix every time! The matrices differs from each other > significantly and there is also difference in number of inliers/outliers! > > Is this a bug? Any one who could provide any suggestion on robust computation > of fundamental matrix is very much appreciated. > > Regards, > Jun I haven't used this particular code, but here are my observations based on another implementation of the algorithm: RANSAC takes random samples of the minimum number of feature point correspondences necessary to compute the fundamental matrix, and computes how many of the entire set of correspondences are consistent with that matrix (the socalled "inliers"). The number of iterations (trials) to be run is continually updated, based on the the number of trials so far, the maximum number of inliers found, total sample size, and the userspecified confidence of the solution's correctness. Because the sampling is random, the number of trails and best solution found will vary from run to run. You could try increasing the confidence threshold to choose more samples, or, if consistency is important, set the random number generator seed before each run. If only a small fraction of the correspondences that you feed the algorithm are correct, RANSAC may have a hard time to find the correct solution, perhaps due in part to epipolar matching ambiguity. Harry Voorhees _________________________________________________________ Harry L. Voorhees, hlv@..., 9784755279 Stellar Science Ltd. Co., http://www.stellarscience.com "Stellar Scientific Software Solutions" 