Design and evaluation of feature detectors

Many applications in both image processing and computational vision rely upon the robust detection of parametric image features and the accurate estimation of their parameters. In this thesis, I address three fundamental questions related to the design and evaluation of parametric feature detectors. Most feature detectors have been designed to detect a single type of feature, more often than not, the step edge. A large number of other features are also of interest. Since the task of designing a feature detector is very time consuming, repeating the design effort for each feature is wasteful. To address this deficiency, in the first part of this thesis I develop an algorithm that takes as input a description of a parametric feature and automatically constructs a detector for it. The development of many feature detectors begins with an ideal model of the feature. Since image data are noisy, feature detectors must actually detect features that are almost, but not quite, ideal. Many existing feature detectors can therefore be regarded as being defined by two components: (1) an ideal feature model, and (2) a function that measures how far the image data may deviate from ideal and still be regarded as the feature. For many detectors, little consideration has been given to the selection of the second of these two components. In the second part of this thesis, I present a method of selecting the deviation function to maximize the performance of the general purpose feature detector described in the first part. Essentially only two methods have actually been used to evaluate feature detectors empirically. The first consists of applying the detectors to a small number of real images and getting a human to evaluate the outputs. The second method involves generating a large number of synthetic images and then computing performance metrics from the outputs using knowledge of the way that the synthetic images were generated. Both of these approaches have their flaws. The first method is subjective. The second method does not use real images. In the third and final part of this thesis, I propose a class of evaluation techniques that use a large number of real images, and yet provide non-subjective performance metrics.