evaluation of accuracy of cvfindhomography

asked 2013-01-30 23:59:44 -0500

pranav gravatar image

updated 2013-01-31 00:24:05 -0500


I am planning an experiment to evaluate accuracy of homography computed using cvfindhomography. I want to compare RANSAC & LMEDS options. I will select N known points(coordinates are known) on a planer screen and will capture its view using a webcam. Then i will select a subset of points from them for computing homography between screen and camera sensor plane. Using this homography i will project all other screen points(which were not used for homography estimation) to camera image plane. Error criteria for homography then will be radial distance of projections using estimated homography and their true position as detected in the undistorted image(assuming corner detector detects corners accurately).

Actually the target is to model the behaviour of error in estimated corner locations so as to determine an error compensation term which can be used to correct the corner locations estimated by homography.

Has anyone tried similar experiment with RANSAC & LMEDS for cvfindhomography or has any suggestion for a better approach to model such error term?

edit retag flag offensive close merge delete