added tests for http://code.opencv.org/issues/4011 and http://code.opencv.org/issues/3057; fixed random subset generation in both methods to increase chance for a good subset