Localized Global Descriptors for Image Retrieval: An Extensive Evaluation on Adaptations to the Simple Model

Abstract

SIMPLE (Searching Images with MPEG-7 (& MPEG-7-like) Powered Localized dEscriptors) is a model that proposes the reuse of well-established global descriptors by localizing their description mechanism on image patches located by local features' detectors. Having displayed impressive retrieval results on two different databases, in this paper we extend the family by replacing the originally picked global descriptors and by applying VLAD, a BOVW alternative, for the vectorization of the features. We re-evaluate all SIMPLE descriptors (original and proposed here) on five benchmarking databases featuring very diverse scenarios, so as to investigate how SIMPLE works with different image retrieval cases, ranging from near duplicate search to visual object similarity. The experimental results show the robustness of the scheme with SIMPLE descriptors presenting stable and high retrieval performances across all tested collections, outperforming not only the methods they originated from, but also some of the best reported state-of-the-art methods as well.