Appearance-based object recognition systems rely on training from imagery, which allows the recognition of objects without requiting a 3D geometric model. It has been little explored whether such systems can be trained from imagery that is unlabeled, and whether they can be trained from imagery that is not trivially segmentable. In this paper we present a method for minimally supervised training of a previously developed recognition system from unlabeled and unsegmented imagery. We show that the system can successfully extend an object representation extracted from one black background image to contain object features extracted from unlabeled cluttered images and can use the extended representation to improve recognition performance on a test set.

Object recognition from a single view fails when the available features are not sufficient to determine the identity of a single object, either because of similarity with another object or because of feature corruption due to clutter and occlusion. Active object recognition systems have addressed this problem successfully, but they require complicated systems with adjustable viewpoints that are not always available. In this paper we investigate the performance gain available by combining the results of a single view object recognition system applied to imagery obtained from multiple fixed cameras. In particular, we address performance in cluttered scenes with varying degrees of information about relative camera pose. We argue that a property common to many computer vision recognition systems, which we term a weak target error, is responsible for two interesting limitations of multi-view performance enhancement: the lack of significant improvement in systems whose single-view performance is weak, and the plateauing of performance improvement as additional multi-view constraints are added.

In this paper we explorer the problem of training a general, 3D abject recognition system from unlabeled imagery. In particular, we attempt to identify critical issues and stumbling blocks associated with minimizing the supervision necessary to train such a system. As class learning seems to be a relatively slow and resource intensive process even for people, we consider approaches and perform experiments that entail on the order of 10/sup 1/5 basic operations, even for relatively small databases. This is the current practical limit of the computation that can be achieved. For experiments, we use a recognition system developed previously.

Appearance-based object recognition systems are currently the most successful approach for dealing with 3D recognition of arbitrary objects in the presence of clutter and occlusion. However, no current system seems directly scalable to human performance levels in this domain. In this paper we describe a series of experiments on a previously described object recognition system that try to see which, if any design axes of such systems hold the greatest potential for improving performance. We look at the potential effect of different design modifications and we conclude that the greatest leverage lies at the level of intermediate feature construction. 1.

In this paper we consider the problem of 3D object recognition and the role that perceptual grouping processes must play. In particular, we argue that reliance on a single level of perceptual grouping is inadequate, since it is responsible for the specific weaknesses of several well-known recognition techniques. Instead, recognition must use a hierarchy of perceptual grouping processes. We describe an appearance-based system that uses four distinct levels of perceptual grouping to represent 3D objects in a form that allows not only recognition, but reasoning about 3D manipulation of a sort that has been supported in the past only by 3D geometric models. The results of the algorithms have been previously reported, and the main contribution of this paper is the development of the perceptual organization hierarchy.

We describe an appearance-based object recognition system using a keyed, multi-level contest representation reminiscent of certain aspects of cubist art. Specifically, we utilize distinctive intermediate-level features in this case automatically extracted 2-D boundary fragments, as keys, which are then verified within a local contest, and assembled within a loose global contest to evoke an overall percept. This system demonstrates good recognition of a variety of 3-D shapes, ranging from sports cars and fighter planes to snakes and lizards with full orthographic invariance. We report the results of large-scale tests, involving over 2000 separate test images, that evaluate performance with increasing number of items in the database, in the presence of clutter, background change, and occlusion, and also the results of some generic classification experiments where the system is tested on objects never previously seen or modeled. To our knowledge, the results we report are the best in the literature for full-sphere tests of general shapes with occlusion and clutter resistance.

This paper presents an approach based on the evolution strategy for autonomous learning of dextrous manipulation primitives with a dextrous robot hand. We use heuristics derived from observations made on human hands to reduce the degrees of freedom of the task and make learning possible. Our system does not rely on simulation; all the experimentation is performed the 16-degree-of-freedom Utah/MIT hand. We present experimental results that show that accurate dextrous manipulation skills can be learned in a period of a few minutes. We also show the application of the learned primitives to perform an assembly task.

We present an experimental evaluation of adaptive and non-adaptive visual servoing in 3, 6 and 12 degrees of freedom (DOF), comparing it to traditional joint feedback control. While the purpose of experiments in most other work has been to show that the particular algorithm presented indeed also works in practice, we do not focus on the algorithm but rather on properties important to visual servoing in general. Our main results are: positioning of a 6 axis PUMA 762 arm is up to 5 times more precise under visual control than under joint control; positioning of a Utah/MIT dextrous hand is better under visual control than under joint control by a factor of 2; and a trust-region-based adaptive visual feedback controller is very robust. For m tracked visual features the algorithm can successfully estimate online the m/spl times/3 (m/spl ges/3) image Jacobian (J) without any prior information, while carrying out a 3 DOF manipulation task. For 6 and higher DOF manipulation, a rough initial estimate of J is beneficial. We also verified that redundant visual information is valuable. Errors due to imprecise tracking and goal specification were reduced as the number of visual features, m, was increased. Furthermore highly redundant systems allow us to detect outliers in the feature vector and deal with partial occlusion.

We propose a behavioral construct called a "virtual tool" that allows many of the advantages of high-degree-of-freedom robotic systems to be realized without excessive computational costs or modeling requirements. The basic idea of a virtual tool is drawn from the observation that the most efficient way of performing a task is generally to use a device that is specially designed for it; this suggests that a way to use redundant degrees of freedom in a system is to use them to customize it so that it becomes, in essence, a special purpose tool for the task at hand. We call the resulting instantiation a virtual tool. Formally, this tailoring takes the form of implementing constraints on the degrees of freedom of the system through lower-level control processes. Remaining degrees of freedom appear as control parameters of the virtual tool, and can be used by higher-level processes, or by a human teleoperator, to control the system. We present experimental results showing the application of the virtual tool approach to telemanipulation using a 22 degree-of-freedom hand-arm system. Our chosen task is a fairly complex assembly task requiring the robot to pick up a small object, transport it, and insert it into a hole with matching shape. Instead of attempting to map the configuration of the teleoperator's hand-arm system to the robot using a data-glove or similar device, we provide basic parameterized grasping and manipulation primitives (or virtual tools) that can be invoked by the teleoperator as he progresses through the stages of the complex task.

The recognition of repetitive movements characteristic of walking people, galloping horses, or flying birds is a routine function of the human visual system. It has been demonstrated that humans can recognize such activity solely on the basis of motion information. We demonstrate a general computational method for recognizing such movements in real image sequences using what is essentially template matching in a motion feature space coupled with a technique for detecting and normalizing periodic activities. This contrasts with earlier model-based approaches for recognizing such activities.

A method of activity detection is described. This technique uses a periodicity measure on gray-level signals extracted along spatio-temporal reference curves. The technique is illustrated using real-world examples of activities. It is shown that the technique robustly detects complex periodic activities, while excluding nonperiodic motion. A technique to recognize these activities using the detection scheme described is proposed. It is not clear how much the periodicity alone is useful for recognition, but the authors believe that the phase information is valuable for activity recognition.>

A method is described for extracting lineal features from an image using extended local information to provide robustness and sensitivity. The method utilizes both gradient magnitude and direction information, and incorporates explicit lineal and end-stop terms. These terms are combined nonlinearly to produce an energy landscape in which local minima correspond to lineal features called sticks that can be represented as line segments. A hill climbing (stick-growing) process is used to find these minima. The method is compared to two others, and found to have improved gap-crossing characteristics.>

A method of visual motion recognition applicable to a range of naturally occurring motions that are characterized by spatial and temporal uniformity is described. The underlying motivation is the observation that, for objects that typically move, it is frequently easier to identify them when they are moving than when they are stationary. Specifically, it is shown that certain statistical spatial and temporal features that can be derived from approximations to the motion field have invariant properties, and can be used to classify regional activities such as windblown trees, ripples on water, or chaotic fluid flow, that are characterized by complex, non-rigid motion. The technique is referred to as temporal texture analysis, in analogy to the techniques developed to classify gray-scale textures. The techniques are demonstrated on a number of real-world image sequences containing complex movement. The work has practical application in monitoring and surveillance, and as a component of a sophisticated visual system.>

Two complementary methods for the detection of moving objects by a moving observer are described. The first (constraint ray filtering) uses a constraint that restricts the projected velocity at any image point to a 1-D locus in velocity space to detect motion inconsistent with a rigid world assumption. The second (animate motion detection) utilizes a constraint on the time-rate-of-change of projected velocity due to smooth observer motion to detect moving objects such as animals and maneuvering vehicles whose projected motion changes rapidly. In both cases, the qualitative nature of the constraints allows the methods to be used with the inexact motion information typically available from real image sequences. Implementations of the methods that run in real time on a parallel pipelined image processing system are described.>

A description is given of the Independent Agents Architecture, designed for building behavior-based robots which make use of sophisticated, computationally expensive sensory modalities. Systems built using this architecture have three layers corresponding to perception, behavior, and motor control. The use of a shared perception layer prevents unnecessary redundancy in the processing of sensor data. Robotic tasks are decomposed into independent stimulus/response behaviors, each of which is implemented by a separate agent. The formalism of stimulus/response spaces is introduced to model these sensorimotor behaviors. This formalism provides a method to analyze when behavioral systems are free from potential action conflicts. In addition, it is demonstrated that it is possible to build a robotic system which performs dynamic tasks through the use of qualitative sensorimotor behaviors instead of mathematical models of the physical world. Juggler, a behavior-based balloon-bouncing robot, is an example of such a system.>

This paper describes QUILT, a prototype geographic information system (GIS) that uses the quadtree data structure as the underlying representation for cartographic data. While QUILT contains many features typically available in a GIS, its primary purpose is to serve as a testbed for the design and testing of new data structures and algorithms for use in computer cartography. Quadtree variants for region, point and line data are implemented using the linear quadtree, organized on disk by a B-tree. QUILT provides a simple attribute attachment system which associates non-spatial data with geographic objects. The user views QUILT as an augmented LISP environment. QUILT's geographic functions include conversion of rasters to and from quadtrees; subset operations to select specified geographic objects; map editing, display, windowing, intersection and union operations; polygon expansion; and computation of geographic object properties such as the centroid, area, perimeter and bounding rectangle for sets of geographic objects.

The use of certain measures of flow field divergence is investigated as a qualitative cue for obstacle avoidance during visual navigation. It is shown that a quantity termed the directional divergence of the 2-D motion field can be used as a reliable indicator of the presence of obstacles in the visual field of an observer undergoing generalized rotational and translational motion. The necessary measurements can be robustly obtained from real image sequences. Experimental results are presented showing that the system responds as expected to divergence in real-world image sequences, and the use of the system to navigate between obstacles is demonstrated.>

The use of certain measures of flow field divergenee is investigated as a qualitative cue for obstacle avoidance during visual navigation. It is shown that a quantity termed the directional divergence of the 2-D motion field can be used as a reliable indicator of the presence of obstacles in the visual field of an observer undergoing generalized rotational and translational motion. Moreover, the necessary measurements can be robustly obtained from real image sequences. A procedure for robustly extracting divergence information from image sequences which can be performed using a highly parallel architecture is described. Experimental results are presented showing that the system responds as expected to divergence in real world image sequences, and the use of the system to navigate between obstacles is demonstrated.

A new method termed population analysis is presented for approximating the distribution of node occupancies in hierarchical data structures which store a variable number of geometric data items per node. The basic idea is to describe a dynamic data structure as a set of populations which are permitted to transform into one another according to certain rules. The transformation rules are used to obtain a set of equations describing a population distribution which is stable under insertion of additional information into the structure. These equations can then be solved, either analytically or numerically, to obtain the population distribution. Hierarchical data structures are modeled by letting each population represent the nodes of a given occupancy. A detailed analysis of quadtree data structures for storing point data is presented, and the results are compared to experimental data. Two phenomena referred to as aging and phasing are defined and shown to account for the differences between the experimental results and those predicted by the model. The population technique is compared with statistical methods of analyzing similar data structures.

A consistent hierarchical data structure for the representation of vector data is presented. It makes use of a concept termed a line segment fragment to prevent data degradation under splitting or clipping of vector primitives. This means that the insertion and subsequent deletion (and vice versa) of a vector leaves the data unchanged. Vectors are represented exactly and not as digital approximations. The data is dynamically organized by use of simple probabilistic splitting and merging rules. The use of the structure for implementing a geographic information system is described. Algorithms for constructing and manipulating the structure are provided. Results of empirical tests comparing the structure to other representations in the literature are given.