Unsupervised Internet-based category learning for object recognition

Abstract

This paper describes a new method for acquiring object categories by searching images on the Internet. An unsupervised method is proposed that, starting from a set of objects extracted from the Internet images automatically fetched for a given category name, selects a subset of objects suitable for building a model of the category. The method is based on repeated k-means clustering. Object relevance scoring is based on properties of the clusters in which they are placed. The approach is evaluated on generic categories (i.e. those usually referenced through common nouns). We demonstrate that the proposed approach significantly improves the quality of the training object collections. © 2013 Springer-Verlag.