Evaluating inverted files for visual compact codes on a large scale

The Bag-of-Words (BoW) or bag-of-features approach is the most widely adopted method for finding images that represent the same object or location given an image as a query and a large set of images as dataset. One of the reasons for the BoW success is that off-the-shelf text search engines can be used to build a distributed and robust system. Very recently, Vector of Locally Aggregated Descriptors (VLAD) have been shown to yield higher effectiveness than BoW. However, VLAD cannot be used in place of BoW on top of text search engines given that it is not a sparse vector of occurrence counts. In this paper, we propose combining a Surrogate Text Representation (SRT) approach with VLAD allowing indexing with traditional text search engines. While previous comparisons between VLAD and BoW were conducted using distinct indexing techniques, we show that the VLAD-SRT combination significantly outperforms BoW on the very same hardware and software infrastructure. The experiments show that VLAD-SRT outperforms BoW and that SRT does not significantly reduce the effectiveness of VLAD.