Fig. 3.

Relative validation set accuracy versus number of training images. We studied the relative validation set accuracy as a function of the number of images in the training set. The training set is randomly chosen as a subset of the original training set at rates of (5%, 10%, …, 100%). Each set includes all the images in the smaller subset. As shown in the figure, the performance does not improve too much after the training set includes more than 3712 images