Experimental parameters and setup

In order to make the training collage easier to construct, all images were re-sized to be the same size (150 by 113). The epitome parameters K,N,T were determined empirically by preliminary trial-and-error experiments. For the image and training collage sizes we used, setting K=15, N=75, and T=1412 seemed to work well in simple trial experiments. For the faces dataset, I trained on 5 positive and 5 negative examples, then tested on the remaining 8 positive and 8 negative examples. For the beaches dataset, I trained on 8 positive and 8 negative examples, then tested on the remaining 5 positive and 5 negative examples. Originally, the classifier performed very poorly (around 30% accuracy on positive test images) on the beaches dataset when using only 5 pairs of training images.