"Tell me more": How semantic technologies can help refining internet image search

Several branches of computer vision heavily rely (but we could even say depend) on the availability of large datasets of labelled images. While such labeling is usually done by hand, a powerful help can be obtained from Internet and its related tools. In this paper we address the problem of automatically generating a set of images representing an object class, given the name of the class. We exploit semantic technologies, such as lexical resources and ontologies, in order to improve the search performances by using a standard web search engine. We will also discuss an application to the automatic building of a training set for a classification framework. Preliminary experiments are provided for 10 classes from the public CalTech256 dataset and results show an average increment in classification accuracy of about 10%. © 2013 ACM.