Improving accessibility of virtual worlds by automatic object labeling

Abstract

User generated virtual worlds, such as Second Life, typically lack accurate metadata for their virtual world objects. This is a significant problem for blind users who rely on textual descriptions in order to access virtual worlds using synthetic speech. In this paper, we consider the problem of automatic object labeling to improve accessibility of virtual worlds for users with disabilities. Taking advantage of the primitivebased representation of virtual world objects in Second Life, we present an approach that leverages histogram-based geometric object representations, machine learning and crowdsourcing to accurately label virtual world objects at a large scale. We report excellent classification results using seven challenging object classes. © 2013 Springer-Verlag.