A comparison of methods for semantic photo annotation suggestion

In this work we describe a general framework for semi-automated semantic digital photo annotation though the use of suggestions. We compare context-based methods with Latent Semantic Indexing, a linear algebra approach to information retrieval. Through experiments on real data sets containing up to 13,705 semantically annotated photos, we show that a carefully chosen combination of context-based methods can not only be efficient, but also extremely effective as well. Furthermore, we propose a new combination of context-based methods that outperforms previous work by up to 19% higher recall while running up to 21 times faster.