A Novel Progressive Transmission in Mobile Visual Search

Abstract

How to reduce the transmission latency is a main concern in the context of Mobile Visual Search (MVS). Transmitting extremely compacted visual descriptor in a progressive manner is the start-of-art solution. In this paper, we present a novel MVS system following the client-server architecture. To reduce the transmission latency, the inquiry image is represented by a set of hash bits, which are then progressively transmitted. In the server side, all images are indexed by their hash bits, similar as the classic Bag-of-Word (BoW) model. Owe to the merit of the proposed system, the IDF weight of the hash bits are encoded into a sparse vector which retained in the mobile client, and provides a transmission order of the inquiry hash bits. The hash bit with lower IDF weight will be more discriminative, which should have higher priority during the transmission. As far as we know, this work is the first one attempting to transmit the hash bits in a proper progressive manner in MVS. Extensive experiments have been done on the public Stanford MVS database, demonstrating that the proposed progressive transmission strategy achieves higher recognition rate compared to other strategies, when delivering the same amount of data.