To evaluate the performance of our method, an image
sequence with 1000 frames was used. In the experiment, we
compared the re-projection error and the number of inliers both
in normal conditions when we could keep the variety of camera
views to a minimum and in scale variation conditions with
dynamic camera view changes. In Fig. 7, it can be seen that our
method, as expected, produces more inliers and less reprojection error than does the GRF with a single sub-region in
most of the time in scale variation conditions. This also means
that our tracking method has less flickering than existing GRF
does.