The authors present a framework to guide development of evidential reasoning in object recognition systems. Principles of evidential reasoning processes for open-world object recognition are proposed and applied to build evidential reasoning capabilities. The principles summarize research and findings by the authors up through the mid-1990s, including seminal results in object-centered computer vision, figure-ground discrimination, and the application of hierarchical Bayesian inference, Bayesian networks, and decision graphs to evidential reasoning for object recognition.

We present the results of an empirical study in which we evaluated cost-sensitive learning algorithms on a rooftop detection task, which is one level of processing in a building detection system. Specifically, we investigated how well machine learning methods generalized to unseen images that differed in location and in aspect. For the purpose of comparison, we included in our evaluation a handcrafted linear classifier, which is the selection heuristic currently used in the building detection system. ROC analysis showed that, when generalizing to unseen images that differed in location and aspect, a naive Bayesian classifier outperformed nearest neighbor and the handcrafted solution.

This paper discusses some representation issues and challenges involved in object recognition. It is intended as a step toward assessing current object representation schemes and proposing design and evaluation criteria for future ones.>

The authors present a new in vivo method to correct the nonlinear, object-shape-dependent and material-dependent spatial distortion in magnetic resonance (MR) images caused by magnetic susceptibility variations. This distortion across the air/tissue interface before and after the correction is quantified using a phantom. The results are compared to the distortion-free computed tomography (CT) images of the same phantom by fusing CT and MR images using fiducials, with a registration accuracy of better than a millimeter. The distortion at the bone/tissue boundary is negligible compared to the typical MRI (MR imaging) resolution of 1 mm, while that at the air/tissue boundary creates displacements of about 2 mm (for G(x) 3.13 mT/m). This is a significant value if MRI is to provide highly accurate geometric measurements, as in the case of target localization for stereotaxic surgery. The correction scheme provides MR images with accuracy similar to that of CT: 1 mm. A new method to estimate the magnetic susceptibility of materials from MR images is presented. The magnetic susceptibility of cortical bone is measured using a SQUID magnetometer, and is found to be -8.86 ppm (with respect to air), which is quite similar to that of tissue (-9 ppm).

The history of the image understanding environment (IUE) project, a five-year program to develop a common software environment for the development of algorithms and application systems, is reviewed. An overview of some of the data structures that are currently evolving as a specification for the IUE is provided. The ultimate goal of the project is to provide the basic data structures and algorithms that are required to carry state-of-the-art research in image understanding.>

Algorithms to detect pairs of edges that could be ends of a straight homogeneous generalized cylinder (SHGC) are proposed. Geometrical constraints for the ends of an SHGC are utilized to group edge elements and edge segments in a complex image. Two methods are investigated. The first algorithm is for a subset of SHGCs for which scaling factors of the cross section at both ends are the same. The second algorithm is for any SHGC. However, a modified version is implemented to reduce computation; given a reference and edge, it finds the edges possibly paired with it. Several examples of ends extracted from real images are reported to show the feasibility of the algorithm.>

It is shown how contact curvature can be determined from a single contact with a cylindrical tactile sensor. When the tactile finger touches an unknown smooth convex surface, contact location, principal curvatures, and normal force are determined from a 4*4 window of strain measurements. Contact properties are determined by a nonlinear model-based inversion from strain measurements back to the contact type. Sensor strains are predicted by convolving the spatial impulse response of the rubber skin with the assumed surface pressure distribution derived from a Hertz contact model. Gradient search finds the parameters of the convex second-order shape and the force that best fit the sensor data. Experiments under laboratory conditions show radius estimation within 10%, orientation within 3%, and subtactel (tactile element) localization to 3% of the element spacing. Using a linearized model, error bounds due to sensor noise on the inversion process are predicted.>

A mobile robot that autonomously functions in a complex and previously unknown indoor environment has been developed. The omnidirectional mobile robot uses stereo vision, odometry, and contact bumpers to instantiate a symbolic world model. Finding stereo correspondences across a single epipolar line is adequate for instantiating the model. Uncertainty in sensor data is represented by a multivariate normal distribution, and uncertainty models for motion and stereo are presented. Uncertainty is reduced by extended Kalman filtering. To execute a high-level command such as 'Enter the second door on the left', a model is instantiated from sensing and motions are planned and executed. Experimental results from the fast, running system are presented.>

The authors develop a color space metric which is useful for computer vision. While there is no shortage of proposed methods for quantifying how the human eye perceives color differences, it is difficult to apply these perceptual metrics directly to the color images sensed by a computer vision system. The metric developed can be applied to images sensed using color filters (e.g. red, green, or blue) as is often done in computer vision. The metric is defined in terms of the spectral characteristics of the filters and camera and accounts for their noise properties. The color distance function is derived as an estimate of the physical distance between normalized spectral power distributions. Components of this distance are weighted to account for sensor noise properties. A color metric has several possible uses in a computer vision system. The usefulness of the color metric is demonstrated by evaluating its performance on regions of real images.>

A description is given of a device for determining curvatures using a cylindrical tactile sensor. The finger touches an unknown convex surface. Principal curvatures, normal force and location are determined from a 4*4 window of strain measurements. Sensor strains are predicted by convolving the spatial impulse response of the rubber skin with the assumed surface pressure distribution derived from a Hertz contact model. Gradient search finds the parameters of the convex second-order shape and the force that best fit the sensor data. Experiments show radius estimation within 10%, orientation within 2.5 degrees, and sub-tactel localization of 3% of the element spacing. Accuracy limits due to sensor noise are derived.>

The authors introduce the concept of an explicit generic building model. This abstract model is valid for a broad class of buildings. It uses mechanisms that have been used to build generic models for other object classes, including machine screws and generalized cylinders. These methods extend the generality of object classes which can be implemented, compared to previous implementations of object classes. Using a combination of epipolar stereo vision from the mobile robot along with monocular vision, the generic model is instantiated. This instantiation can be used for robot navigation.>

A new version of the AL language system has been implemented, with emphasis on the specifications for an object's motion in operational space under trajectory planning, and flexible grasping. This paper discusses specifications of object geometry, Cartesian trajectory, control frame calculation from the trajectory and geometry, and motion constraints with respect to the control frame. A simple inverse of a Jacobian matrix linking the control frame to joint space variables is also presented.

A mobile robot architecture must include sensing, planning, and locomotion which are tied together by a model or map of the world based on sensor information, apriori knowledge and generic models. The architecture of a Stanford's autonomous mobile robot is described including its distributed computing system, locomotion, and sensing. Additionally, some of the issues in the representation of a world model are explored. Sensor models are used to update the world model in a uniform manner, and uncertainty reduction is discussed.

An edge in an image corresponds to a discontinuity in the intensity surface of the underlying scene. It can be approximated by a piecewise straight curve composed of edgels, i.e., short, linear edgeelements, each characterized by a direction and a position. The approach to edgel-detection here, is to fit a series of one-dimensional surfaces to each window (kernel of the operator) and accept the surfacedescription which is adequate in the least squares sense and has the fewest parameters. (A one-dimensional surface is one which is constant along some direction.) The tanh is an adequate basis for the step-edge and its combinations are adequate for the roof-edge and the line-edge. The proposed method of step-edgel detection is robust with respect to noise; for (step-size/noise) > 2.5, it has subpixel position localization (position < 3) and an angular localization better than 100; further, it is designed to be insensitive to smooth shading. These results are demonstrated by some simple analysis, statistical data, and edgel-images. Also included is a comparison of performance on a real image, with a typical operator (Difference-of-Gaussians). The results indicate that the proposed operator is superior with respect to detection, localization, and resolution.

A class of inferences is described which allows the recovery of three-dimensional structures from the two-dimensional curves in an image. Unlike most previous methods, these inferences do not require restrictive assumptions or prior knowledge regarding the scene. They are based on the assumption that the camera viewpoint and the positions of the illumination sources are independent of the objects in the scene. From these independence assumptions, it can be shown that many potential interpretations of image curves are highly improbable. By eliminating these improbable interpretations it is possible to segment the image into sets of related image features and derive many three-space relations.

A ranging system, consisting of a laser, computer-controlled optical deflection assembly, and TV camera, obtains three-dimensional images of curved solid objects. The object is segmented into parts by grouping parallel traces obtained from the ranging system. Making use of the property of generalized translational invariance, the parts are described in terms of generalized cylinders, consisting of a space curve, or axis, and a circular cross section function on this axis.