In model based recognition, the goal is to locate an instance of one or more known objects in an image. The problem is compounded in real images by the presence of clutter, occlusion, and sensor error, which can lead to "false negatives", failures to recognize the presence of the object, and "false positives", in which the algorithm incorrectly identifies an occurrence of the object. The probability of either event is affected by parameters within the recognition algorithm, which are almost always chosen in an ad-hoc fashion. The effect of the parameter values on the likelihood that the recognition algorithm will make a mistake are usually not understood explicitly. To address the problem, we explicitly model the noise that occurs in the image. In a typical recognition algorithm, hypotheses about the position of the object are tested against the evidence in the image, and an overall score is assigned to each hypothesis. We use a statistical model to determine what score a correct or incorrect hypothesis is likely to have, and use standard binary hypothesis testing techniques to distinguish correct from incorrect hypotheses. Using this approach, we can compare algorithms and noise models, and automatically choose values for internal system thresholds to minimize the probability of making a mistake.

The probability of false positives and negatives is derived as a function of the number of model features, image features, and occlusion, under the assumption of 2D Gaussian noise and a particular method of evidence accumulation. No assumptions are made about prior distributions on the model space, nor is even the presence of the model assumed. The results are presented in the form of ROC (receiver-operating characteristic) curves, from which several results can be extracted. They demonstrate that the 2D Gaussian error model has better performance than that of the bounded uniform model for the same level of occlusion and clutter. They also directly indicate the optimal performance that can be achieved for a given clutter and occlusion rate and how to choose the thresholds to achieve the desired rates. These ROC curves are verified in the domain of simulated images.>

The author shows how it is possible for a mobile robot to exploit the visual information obtained by scanning a room to determine its size and shape, and to orient itself continually within it. The equipment used is a very simple camera setup whose detailed initial configuration is not known but can be deduced as the algorithm runs. The approach does not require any special environment, nor is it sensitive to changes in the physical aspect of the room being inspected such as moved furniture or roaming people. The long-term goal of the project is for the robot to use the information thus acquired in order to build maps of its environment, presumed to be a single floor of an office building, and to localize itself within this framework.>