Entropy-of-likelihood feature point selection for image correspondence

In this thesis, we present a general, non-subjective method of selecting informative feature points for the task of image correspondence, the entropy-of-likelihood (EOL). Feature point selection requires identifying the points in one image that can be most reliably identified and matched in a second image. Most feature point selection methods identify points that are interesting according to subjective notions as to which points are best for matching, such as points with a high degree of edge density, nostrils in face images, etc. The EOL method differentiates itself from the majority of feature point selection methods in that points are chosen by explicitly evaluating points according to their potential to result in fast, unique correspondence, given the particular model used for correspondence and images to be matched. We describe the EOL feature point selection within the framework of the Bayesian Markov random field (MRF) correspondence model, where the degree of feature point information is encoded by the entropy of the Bayesian likelihood term. We propose that feature selection according to minimum entropy-of-likelihood (EOL) is less likely to lead to correspondence ambiguity, thus improving the optimization process in terms of speed and quality of solution. Experimental results demonstrate the ability of the EOL to select optimal feature points in a wide variety of image contexts, such as objects, faces, aerial photographs, etc. Correspondence trials comparing EOL feature point selection with the well-known Kanade-Lucas-Tomasi (KLT) method reveal that EOL feature points lead to correspondence that is significantly faster and less likely than KLT to result in sub-optimal, locally maximal solutions. In addition, ground truth comparisons show that EOL feature points result in a lower residual error.