This program computes the trilinear tensor using a non-linear 6 point algorithm of Torre. In the same way as compfund the tensor is refit to the support correspondences at the end of the random sampling process, but only if this decreases the average error (which is often the case).
comptensor -f matchesfile [-n numberOfRandomSamples] [-count] [-l2]
[-a affine] [-m no refit] [-s stop early count] [-d distance for inliers]
[-i imagefile1] [-j imagefile2] [-k imagefile3]
-f: the triple correspondence match file used to compute the fundamental matrix.
This is a mandatory input.
-n: the number of random samples to perform.
-count: use count scoring (score by counting the number of supporting points).
-l2: use l2 scoring (score supporting points based on the inlier distance).
-a: compute only the affine tensor.
-m: do not perform a refit at the end.
-d: maximum allowable distance to be considered an inlier
-i: 1st image file for drawing translation vectors into
-j: 2nd image file for drawing translation vectors into
-k: 3rd image file for drawing translation vectors into
-s: success probability, stop sampling when this probability of success has been reached
The s flag substantially reduces the required number of random samples by stopping the process when this probability of success has been reached.
The inlier distance for a correspondence triple (three points) is the average error of the tensor transfer. In other words, remove one point, compute the predicted point in that image using the tensor, and compute the distance from this predicted point to the actual point. The average of these three distances is the inlier distance for that correspondence.
count style scoring
1.0 pixels distance to be considered an inlier
NO refit at the end
Success probability: 99.999 percent
Default algorithm: Exact with 6 matches
A file with the same name as the input file is produced with the extension TM. This holds the nine values for the fundamental matrix. E.g. X.matches.filtered à X.matches.filtered.TM