George C. Stockman

Object recognition and localization via pose clustering

The general paradigm of pose clustering is discussed and compared to other techniques applicable to the problem of object detection. Pose clustering is also called hypothesis accumulation and generalized Hough transform and is characterized by a “parallel” accumulation of low level evidence followed by a maxima or clustering step which selects pose hypotheses with strong support from the set of evidence. Examples are given showing the use of pose clustering in both 2D and 3D problems. Experiments show that the positional accuracy of points placed in the data space by a model pose obtained via clustering is comparable to the positional accuracy of the sensed data from which pose candidates are computed. A specific sensing system is described which yields an accuracy of a few millimeters. Complexity of the pose clustering approach relative to alternative approaches is discussed with reference to conventional computers and massively parallel computers. It is conjectured that the pose clustering approach can produce superior results in real time on a massively parallel machine.

A Region-Based Approach to Digital Image Registration with Subpixel Accuracy

Automatic registration of images with translational, rotational, and scaling differences is discussed. To register two images from the same scene, first, the images are segmented and closedboundary regions in the images are extracted. Next, centers of gravity of closed-boundary regions are taken as control points and correspondence is established between the control points. Using this correspondence, the original images are then revisited and the segmentation process is refined in such a way that the obtained corresponding regions become optimally similar. This enables determination of centers of gravity of the regions up to subpixel accuracy. Finally, registration parameters are determined by the least squares error criterion.

A minimax algorithm better than alpha-beta?

Abstract An algorithm based on state space search is introduced for computing the minimax value of game trees. The new algorithm SSS∗ is shown to be more efficient than α-s in the sense that SSS∗ never evaluates a node that α-s can ignore. Moreover, for practical distributions of tip node values, SSS∗ can expect to do strictly better than α-s in terms of average number of nodes explored. In order to be more informed than α-s, SSS∗ sinks paths in parallel across the full breadth of the game tree. The penalty for maintaining these alternate search paths is a large increase in storage requirement relative to α-s. Some execution time data is given which indicates that in some cases the tradeoff of storage for execution time may be favorable to SSS∗.

3-D surface solution using structured light and constraint propagation

A method for 3-D surface measurement using a projected grid of light is presented. The grid line identification problem is solved using very general constraints. At the same time, 3-D surface patch solutions are developed. From the general constraints of set of geometric and topological rules are derived that are effectively applied in the computation of grid labels and hence 3-D surface solutions. A set of five algorithms has been applied on five real scenes consisting of multiple objects of arbitrary shapes. The results show that globally consistent surface solutions can be obtained rapidly with good accuracy using a single image. A small degree of ambiguity remains, but can be further reduced or removed using increased knowledge. >

Point pattern matching using convex hull edges

Algorithms for matching two sets of points in a plane are given. These algorithms search in the parameter space and find the transformation parameters that can match the most points in two sets. Since an exhaustive search for the best parameters is not affordable as the number of points in the sets become large, a subset selection method is given in order to reduce the search domain. Subsets are chosen as points on the boundary of the convex hulls of the sets. The algorithms are tested on generated and real data, and their performance is compared.

Detection of Anchor Points for 3D Face Veri.cation

This paper outlines methods to detect key anchor points in 3D face scanner data. These anchor points can be used to estimate the pose and then match the test image to a 3D face model. We present two algorithms for detecting face anchor points in the context of face verification; One for frontal images and one for arbitrary pose. We achieve 99% success in finding anchor points in frontal images and 86% success in scans with large variations in pose and changes in expression. These results demonstrate the challenges in 3D face recognition under arbitrary pose and expression. We are currently working on robust ?tting algorithms to localize more precisely the anchor points for arbitrary pose images.

Controlling a computer via facial aspect

Control of a computer workstation via face position and facial gesturing would be an important advance for people with hand or body disabilities as well as for all users. Steps toward realization of such a system are reported here. A computer system has been developed to track the eyes and the nose of a subject and to compute the direction of the face. Face direction and movement is then used to control the cursor. Test results show that the resulting system is usable, although several improvements are needed. >

Surface orientation from a projected grid

Two simple methods are given for obtaining the surface shape using a projected grid. After the camera is calibrated to the 3-D workspace, the only input date needed for the computation of surface normals are grid intersect points in a single 2-D image. The first method performs nonlinear computations based on the distortion of the lengths of the grid edges and does not require a full calibration matrix. The second method requires that a full parallel projection model of the imaging is available, which enables it to compute 3-D normals using simple linear computations. The linear method performed better overall in the experiments, but both methods produced normals within 4-8 degrees of known 3-D directions. These methods appear to be superior to methods based on shape-from-shading because the results are comparable, yet the equipment setup is simpler and the processing is not very sensitive to object reflectance. >

Determining pose of 3D objects with curved surfaces

A method is presented for computing the pose of rigid 3D objects with arbitrary curved surfaces. Given an input image and a candidate object model and aspect, the method will verify whether or not the object is present and if so, report pose parameters. The curvature method of Bash and Ullman is used to model points on the object rim, while stereo matching is used for internal edge points. The model allows an object edge-map to be predicted from pose parameters. Pose is computed via an iterative search for the best pose parameters. Heuristics are used so that matching can succeed in the presence of occlusion and artifact and without resetting to use of corresponding salient feature points. Bench tests and simulations show that the method almost always converges to ground truth pose parameters for a variety of objects and for a broad set of starting parameters in the same aspect.

3D Free-Form Object Recognition Using Indexing by Contour Features

We address the problem of recognizing free-form 3D objects from a single 2D intensity image. A model-based solution within the alignment paradigm is presented which involves three major schemes?modeling, matching, and indexing. The modeling scheme constructs a set of model aspects which can predict the object contour as seen from any viewpoint. The matching scheme aligns the edgemap of a candidate model to the observed edgemap using an initial approximate pose. The major contribution of this paper involves the indexing scheme and its integration with modeling and matching to perform recognition. Indexing generates hypotheses specifying both candidate model aspects and approximate pose and scale. Hypotheses are ordered by likelihood based on prior knowledge of pre-stored models and the visual evidence from the observed objects. A prototype implementation has been tested in recognition and localization experiments with a database containing 658 model aspects from twenty 3D objects and eighty 2D objects. Bench tests and simulations show that many kinds of objects can be handled accurately and efficiently even in cluttered scenes. We conclude that the proposed recognition-by-alignment paradigm is a viable approach to many 3D object recognition problems.