CSDD Features: Center-Surround Distribution Distance for Feature Extraction and Matching

Abstract

We present an interest region operator and feature descriptor called Center-Surround Distribution Distance (CSDD) that is based on comparing feature distributions between a central foreground region and a surrounding ring of background pixels. In addition to finding the usual light(dark) blobs surrounded by a dark(light) background, CSDD also detects blobs with arbitrary color distribution that “stand out” perceptually because they look different from the background. A proof-of-concept implementation using an isotropic scale-space extracts feature descriptors that are invariant to image rotation and covariant with change of scale. Detection repeatability is evaluated and compared with other state-of-the-art approaches using a standard dataset, while use of CSDD features for image registration is demonstrated within a RANSAC procedure for affine image matching.