These plots are arranged similar to Figure 13. The SVM with the histogram intersection kernel always produces 100% training rate, so we do not show the training rate in these plots. In contrast to the aLDA model, adding more features always seem to improve the performance of the SVM model. The last three plots show features sets where one feature has been removed. Color and edge related features, namely, curvature, edge-ribbon, and edge-slice, turn out to be more useful than SIFT for FMD images.