Skip to main content
. 2016 Sep 24;122(1):17–33. doi: 10.1007/s11263-016-0950-1

Fig. 5.

Fig. 5

Comparison between GN-DPMs and AAMs. Both algorithms were initialized using the ground truth landmark locations (left) and the ground truth after a relatively large perturbation of the first shape parameter (right). The average (normalized) pt-pt Euclidean error versus fraction of images is plotted. Clearly GN-DPMs are easier to optimize