In all cases: You need at least 4 point correspondences to estimate a perspective tranformation. You can use more than four correspondences and estimate a transformation by least-mean/least-median-of-squares. However, RANSAC (this particular version in OpenCV, not in general) always uses four correspondences and scores the transformation by the size of the concensus set.
In case of RANSAC you therefore need > 4 point correspodences to find the concensus set. Otherwise there are no correspondences besides these that have been used to estimate the transformation.
I think this is not exactly written in the documentation, but it is mentioned everywhere in RANSAC-related literature. E.g. see Hartley and Zisserman, Multiple View Geometry in Computer Vision for details.
In all cases: You need at least 4 point correspondences to estimate a perspective tranformation. You can use more than four correspondences and estimate a transformation by least-mean/least-median-of-squares. However, RANSAC (this particular version in OpenCV, not in general) always uses four correspondences and scores the transformation by the size of the concensus set.
In case of RANSAC you therefore need > 4 point correspodences to find the concensus set. Otherwise there are no correspondences besides these that have been used to estimate the transformation.
I think this is not exactly written explicitly mentioned in the documentation, documentation but it is mentioned described everywhere in RANSAC-related literature. E.g. see Hartley and Zisserman, Multiple View Geometry in Computer Vision for details.
