Hybrid-Indexing Multi-type Features for Large-Scale Image Search

Abstract

Indexing local features with a vocabulary tree and indexing holistic features by compact hashing codes are two successful but separated lines of research. Both of the two indexing models are suited for specific features and are limited to certain scenarios like partial-duplicate search and similar image search, respectively. To conquer such limitations, we propose a novel hybrid-indexing strategy, which incorporates multiple similarity metrics into one inverted index file during off-line indexing. Hybrid-Indexing only requires the Bag-of-visual Words (BoWs) model as input for online query, but could obtain more satisfying retrieval results because the index file conveys hybrid similarities among images. Moreover, hybrid-indexing does not degrade the efficiency of classic BoWs based image search. Experiments on several public datasets manifest the effectiveness and efficiency of our proposed method.