Image retrieval using accurate approximated inverse document frequency of geometry-preserving visual phrases

The most popular approach for large scale image retrieval is to represent images using the bag-of-visual-word (BoV) model. Based on the typical BoV zhang et al. introduce the idea of geometry-preserving visual phrases (GVP) to encode spatial information [1]. Since GVP can only be generated in searching step, it´s impractical to know the inverse document frequency (idf) of GVP in advance. Zhang et al. define the idf of a GVP as the summation of idf weights of visual words in the GVP. But, this kind of approximation is prone to give a much larger value than its real one. In this paper, we propose to use the smallest idf of visual words in a GVP to approximate the idf of the GVP, while keep the efficiency of GVP searching process. Experiments on Oxford 5K and MIR FLICKER 1M datasets show that our approach can achieve better performance compared with GVP.