Semantic visual templates: linking visual features to semantics

Abstract

The rapid growth of visual data over the last few years has lead to many schemes for retrieving such data. With content-based systems today, there exists a significant gap between the user's information needs and what the systems can deliver. We propose to bridge this gap, by introducing the novel idea of semantic visual templates (SVT). Each template represents a personalized view of concepts (e.g. slalom, meetings, sunsets, etc.). The SVT is represented using a set of successful queries, which are generated by a two-way interaction between the user and the system. We have developed algorithms that interact with the user and converge upon a small set of exemplar queries that maximize recall. The SVTs emphasize intuitive models that allow for easy manipulation and queries to be composited. The resulting system performs well, for example with small number of queries in the "sunset" template, we are able to achieve 50% recall and 24% precision over a large unannotated database.