Experimental procedure

In order to test the effectiveness of this classification method, I created 2 datasets each consisting of 13 positive example images and 13 negative example images. For one dataset, the 13 positive example images were close-up images of a single human face, and for the other the 13 example images were all pictures taken at beaches. For both datasets, the negative examples were a haphazard jumble of images that did not fit the description of the positive dataset. All images were obtained from either Google Image Search or Flickr.

To do a single experimental run, I randomly selected an equal number of positive and negative images to be training examples and used them to construct the training collage. The remaining example images were then classified using the discriminative patches extracted from the training collage. This procedure was repeated 3 times for each of the 2 datasets. Running the experiments took between 12 and 13 hours for the faces dataset and about 15 hours for the beaches dataset. The difference is due to the fact that I used more images for training on the beaches dataset, since the classifier performed much worse on the beaches set than on the faces set when using the same number of training examples.