Pattern Recognition from One Example by Chopping

We investigate the learning of the appearance of an object from a single image of it. Instead of using a large number of pictures of an object to be recognized, we use pictures of other objects to learn invariance to noise and variations in pose and illumination. This acquired knowledge is then used to predict if two images of objects unseen during training actually display the same object. We propose a generic scheme called {\it chopping} to address this task. Using a fast learner, we build hundreds of arbitrary binary splits of the image space designed to assign the same label to all the training images of any given object. Predictors of these splits are combined with a Bayesian rule into a posterior probability of similarity. Experiments with the COIL-100 database and with a database of $150$ degraded \LaTeX{} symbols compare our method to a classical learning method with several examples of the positive class and to direct learning of the similarity.

Related material