Boosting VLAD with Weighted Fusion of Local Descriptors

Vector of locally aggregated descriptors (VLAD) is a popular image encoding method for image retrieval. This paper proposes a novel framework to boost VLAD with weighted fusion of local descriptors for discriminative image representation. Due to the fact that most VLAD-based methods generally only use detected SIFT descriptor and contain limited content information, in which the representation ability is deteriorated. In order to obtain a preferable image representation, our approach fuses Dense SIFT and detected SIFT descriptor in the aggregation of local descriptors. Besides, we assign each detected SIFT a weight that measured by saliency analysis to make the salient descriptor with a relatively high importance. In this way, the proposed method can include sufficient image content information and highlight the important image regions. Experiments on image retrieval tasks demonstrate that our approach outperforms previous VLAD-based methods.