Improving classification boundaries by exemplar generation for visual pattern discrimination

In many applications for visual pattern discrimination, a major drawback is insufficient training data. Often the data contains too few example images, and those images are not distributed along the boundary between the alternative classifications. In this paper we present an approach that develops realistic synthetic data along the boundary between two different discrimination classes, where exemplars are needed the most. An application of this technique to a real life object recognition problem shows a performance comparable to that of a human expert, and far better than a network trained only on the available real data. Furthermore, the results are considerably better than those obtained using non-network discriminators such as Euclidean.