BSIFT: Boosting SIFT using principal component analysis

Feature descriptors usually have high dimensionality to efficiently represent key points. Finding matches between large sets of descriptors is a basic step in many applications in computer vision and image processing. When the number of descriptors is large, detection of corresponding points can be extremely time-consuming. The goal of this paper is reducing the computational cost in the matching stage especially for SIFT descriptor. We apply the principal components analysis (PCA) on two sets of SIFT features of images and find a coarse matching between points. Then, the Kullback-Leibler (KL) divergence similarity score is used to improve the matching accuracy. Experimental results show that our proposed technique can reduce the dimension of SIFT and the related matching cost with approximately the same average precision compared to the conventional approach.