Evaluación de una Estrategia de Expansión Local Conservadora en Recuperación de Información Visual

Abstract

In this paper we compare two query expansion methods in the Visual Information Retrieval (VIR) area: Probabilistic Relevance Feedback (PRF) and Local Context Analysis (LCA). The main difference observed between these methods is that while PRF assumes that annotations related to top-ranked images are relevant, LCA avoids to include terms from top-ranked non relevant images of the ranking using an heuristic based on coocurrence. The experiment results show us that LCA increases its precision over PRF for those rankings with lowest precision. Thus, LCA demonstrates to be specially suitable for low precision rankings as the ones returned by the VIR systems based on the content of the image. Indeed, our multimodal LCA variation is the only one local expansion strategy which do not hurt the diversity of the results and the one which reach our best precision results with the ImageCLEFPhoto 2008 task query set – 4° MAP and 5° P20 within the 1039 automatics runs submitted by the participants –.