I. Introduction
Local invariant feature detection [1]–[8] is an important step in a number of applications such as wide baseline matching, object and image retrieval, tracking, recognition, image registration and 3D reconstruction. The classical process to obtain the features consists in detecting a specific class of interest points, such as corners, together with an associated scale generally obtained from a scale-space. Typical examples of such key locations are the local extrema of the result of difference of Gaussians (DoG) applied in scale-space to a series of smoothed and resampled images. Several crucial invariance properties are required for using such points in applications, such as invariance to image translation, scaling, and rotation, to illumination changes or to local geometric distortion.