Towards Unsupervised Discovery of Visual Categories

Recently, many approaches have been proposed for visual object category detection. They vary greatly in terms of how much supervision is needed. High performance object detection methods tend to be trained in a supervised manner from relatively clean data. In order to deal with a large number of object classes and large amounts of training data, there is a clear desire to use as little supervision as possible. This paper proposes a new approach for unsupervised learning of visual categories based on a scheme to detect reoccurring structure in sets of images. The approach finds the locations as well as the scales of such reoccurring structures in an unsupervised manner. In the experiments those reoccurring structures correspond to object categories which can be used to directly learn object category models. Experimental results show the effectiveness of the new approach and compare the performance to previous fully-supervised methods.