Object retrival based on visual word pairs

In object retrieval method based on bag-of-features, local regions of images are characterized using high dimensional descriptors. These descriptors are hierarchically quantized into “visual words” to represent images. One problem of the quantization step is that it reduces the discriminative power of the local descriptors. To address this problem, “descriptor-space soft assignment” mechanism is used to collect the information lost in the quantization step. However, this mechanism also introduces noises, which decreases the precision of image retrieval. In this paper we use two SIFT descriptors, the coarse descriptor and the refined descriptor, to describe an interest point. The experiments show that this approach can efficiently reduce wrong matches caused by descriptor-space soft assignment, and improve the overall performance of an image retrieval system.