MI-SIFT: Mirror and inversion invariant generalization for SIFT descriptor

Abstract

The best known Scale-Invariant Feature Transform (SIFT) shows its superior performance in a variety of image processing tasks due to its distinctiveness, invariance to scale, rotation and local geometric distortion. Despite its remarkable performance, SIFT is not invariant to mirror images and grayscale-inverted images. This paper proposes an improved SIFT descriptor named MI-SIFT which keeps the advantages of the standard SIFT and is additionally invariant to mirror images and grayscale-inverted images. MI-SIFT is achieved by combining SIFT histogram bins in an elegant way at slight expense of dis-tinctiveness. Most importantly, MI-SIFT can be applied to mirror-like images and inversion-like images which are abundant in real world. Experiments show that MI-SIFT outperforms the standard SIFT on mirror-like and inversionlike images while achieve comparable performance on other images. Copyright © 2010 ACM.