Feature selection and aggregation based on rank measures

Abstract

Despite the significant recent advances in feature extraction methods based on deep learning approaches, effectively retrieving images still remains a challenge in various scenarios. Such complexity is mainly associated with the diverse aspects involved in visual perception, which usually can not be encoded by a single feature. Given the myriad of available features, several fusion approaches have been investigated aiming at producing more effective retrieval results. Despite the success of fusion approaches, another crucial task consists in defining what features to combine. Selecting features (or rankers) capable of producing an effective retrieval result consists of a difficult task, especially in unsupervised scenarios, where no information about the effectiveness of the feature is available. This project aims to address this problem by investigating an unsupervised selection framework based on ranking information. The proposed methodology aims to investigate the use of recently proposed rank correlation measures for analyzing complementary information and recent effectiveness estimation measures to identify the most effective combinations. (AU)