Object Recognition by Parts

Abstract

Object recognition is one of the major challenges in computer vision and a vast number of approaches have been proposed. One approach to this problem is to try to recognize an object by building feature detectors for its various parts, and then checking that the parts lie the correct spatial relationships. A key advantage of the recognition-by-parts strategy is that it is robust to problems of occlusion that bedevil strategies based on observing the whole object. In this thesis we examine feature detectors for images of 3-D objects which use a m x m window of the image as the input. A number of feature detectors including multiple logistic regression, linear subspace models, k-nearest-neighbours and different types of artificial neural networks are investigated. The performance of these classifiers has been assessed using both representative test sets and receiver-operating-characteristic (ROC) curves.