Unsupervised Learning of Visual Object Recognition Models

Object recognition from images is traditionally based on a large training set of previously annotated images which is impractical for some applications. Also, most methods use only local or global features. Due to the nature of objects some features are better suited for some objects, so researchers have recently combined both types of features to improve the recognition performance. This approach, however, is not sufficient for the recognition of generic objects which can take a wide variety of appearances. In this paper, we propose a novel object recognition system that: (i) uses a small set of images obtained from the Web, (ii) induces a set of models for each object to deal with polymorphism, and (iii) optimizes the contribution of local and global features to deal with different types of objects. We performed tests with both generic and specific objects, and compared the proposed approach against base classifiers and state-of-the-art systems with very promising results.