Selective Aggregated Descriptors for Robust Mobile Image Retrieval

Towards low latency query transmission via wireless link, methods have been proposed to extract compact visual descriptors on mobile device and then send these descriptors to the server at low bit rates in recent mobile image retrieval systems. The drawback is that such on-device feature extraction demands heavy computational cost and large memory space. An alternate approach is to directly transmit low quality JPEG compressed query images to the server, but the lossy compression results in compression artifacts, which subsequently degrade feature discriminability and deteriorate the retrieval performance. In this paper, we present selective aggregated descriptors to address this problem of mobile image retrieval on low quality query images. The proposed mechanism of selective aggregation largely reduces the negative impact of noisy features caused by compression artifacts, enabling both low latency query transmission from mobile device and effective image retrieval on the server end. In addition, the proposed method allows fast descriptor matching and less storage of visual descriptors for large database. Extensive experiments on benchmark datasets have shown the consistent superior performances of the proposed approach over the state-of-the-art.