An Affine Invariant Interest Point And Region Detector Based On Gabor Filters

This paper presents a novel approach for interest point and region detection which is invariant to affine transformations. Such transformations introduce significant changes in the point location as well as in the scale and the shape of the neighborhood of an interest point. Our approach allows to solve for these problems simultaneously. The approach is based on three key ideas: 1) Interest points can be extracted based on local maxima of the normalized local energy maps. 2) Local extrema over scale of the normalized energy function indicate the presence of characteristic local structures. 3) The maximum response along all the orientations indicates the principle orientation of the local structure. We first extract interest points at multi-scales from the local energy map constructed by Gabor filter responses, and then select points at which a local measure is maximal over scales. This allows a selection of distinctive points for which the characteristic scale is known. We then estimate the principle orientation through the orientational responses of Gabor filters and extend the detector to affine invariance by estimating the affine shape of a point neighborhood. The characteristic scale and the affine shape of neighborhood determine an affine invariant region for each point. Experimental results with synthetic images and natural images show the affine invariance performance of our approach. Comparative evaluation using the repeatability criteria demonstrates the comparable performance in the presence of large viewpoint changes.