Machine Recognition of Objects

Definition Machine recognition of objects is the task of locating and recognizing a given object in an image and consists of the following steps: object detection, feature extraction, and recognition. Background Early computer vision recognition schemes focused primarily on the recognition of rigid three-dimensional (3D) objects, such as machine parts, tools, and cars. This is a challenging problem because the same object can have markedly different appearances when viewed from different directions. It proved possible to deal successfully with this difficulty by using detailed 3D models of the viewed objects, which were compared with the projected 2D image (e.g., [14, 18, 33]). Over the last decade or so, computational models have made significant progress in the task of recognizing natural object categories under realistic, relatively uncon-strained viewing conditions. Within object recognition, it is common to distinguish two main tasks: identification , for instance, recognizing a specific face among other faces, and categorization, for example, recognizing a car among other object classes. We will discuss both of these tasks below and use " recognition " to include both. The qualitative improvement in the performance of recognition models can be attributed to three main components. The first is the use of extensive learning in constructing recognition models. In this framework, rather than specifying a particular model, the scheme starts with a large family of possible models and uses observed examples to guide the construction of a specific model which is best suited to the observed data. The second component was the development of new forms of object representation for the purpose of categorization, based on both computational considerations and guidelines from known properties of the visual cortex. These two components, representation and learning, are interrelated: initially, the class representation provides a family of plausible models, and effective learning methods are then used to construct a particular model for a novel class such as " dog " or " airplane " based on observed examples. The third component was the use of new statistical learning techniques, such as regularization classifiers (SVM and others) and Bayesian inference (such as graphi-cal models). We next discuss each of these advances in more detail. M 470 Machine Recognition of Objects classification problem from examples, rather than focusing on the classifier design. This was a marked departure from the dominant practices at the time: instead of an expert program with a predetermined set of logical rules, the appropriate …