Scale Invariant Feature Transform with Irregular Orientation Histogram Binning

The SIFT (Scale Invariant Feature Transform) descriptor is a widely used method for matching image features. However, perfect scale invariance can not be achieved in practice because of sampling artefacts, noise in the image data, and the fact that the computational effort limits the number of analyzed scale space images. In this paper we propose a modification of the descriptor’s regular grid of orientation histogram bins to an irregular grid. The irregular grid approach reduces the negative effect of scale error and significantly increases the matching precision for image features. Results with a standard data set are presented that show that the irregular grid approach outperforms the original SIFT descriptor and other state-of-the-art extentions.