An investigation into two major potential sources of distortion in stereoscopic displays, namely magnification and vertical disparity, is presented. It is shown that there exist vertical as well as horizontal parallaxes in a 3-D image generated by a pair of TV cameras which are spatially rotated and spaced apart. This could well limit the potential usefulness of stereoscopic television systems, as it unfavorably affects the observer's ability to fuse the corresponding left and right images. An expression for disparity in the y-direction is derived which shows its direct relation to the rotation angle of the TV cameras from the parallel position. Display distortion due to the difference between the depth magnification and the magnification of the xy-plane is also examined. This class of distortion detracts from the realism of the display and hence has undesired effects in tasks involving visual inspection. The equations relating these magnifications to various parameters of the stereoscopic television system are given and the condition under which the two are identical is discussed.

A new signal processing approach toward a multiview 3-D system is described. Stereo image signal processing techniques are used to construct images of virtual cameras (i.e., views from new perspectives) at intermediate positions between real cameras. Compared with conventional systems where a large number of pictures from different viewpoints are required, this allows a dramatic reduction of recording and transmission expenditures.

We assume a left handed coordinate system with interocular distance e and viewer distance to the stereo window d < 0. Using the standard off-axis perspective projection to compute homologous points, the horizontal parallax of a point with coordinates (x, y, z) becomes ez/(z - d). Given a finite set of points D equals {(xj, yj, zj), 1 j in a z-buffer after applying hidden surface elimination. In the first case the solution is based on the root of a quadratic equation. In the second case we approximate the translated parallax function zj + v)/(zj + v - d) by the first term in its series expansion, (zj + v)/d , and use it to analyze the behavior of the sum near the minimum. If we suitably restrict the distance between the closet point and furthest point from the viewer, we argue that the minimum occurs to the right of max {d-zi} and at a root -zj of a parallax function. The root can be located by determining the index i where the function changes sign and evaluating the absolute parallax sum at values -zj where j is close to i.

Stereoscopic computer graphics images, if not properly composed, can be difficult to view. Interaxial separation may be set too low or too high, resulting in an inadequate or excessive stereoscopic effect. Or, the world-coordinate distance from the centers of projection to the plane of zero parallax may be too large or small, leaving the user with image components that seem misaligned and are difficult to fuse into a three-dimensional image. Software designers typically approach this problem by allowing the user to interactively alter two stereoscopic display parameters. However, the user often has difficulty adjusting the parameters of a practically unviewable initial image. Furthermore, changes in the graphics image may require frequent adjustment of stereoscopic parameters. For a good-looking stereoscopic computer graphics image, both negative and positive parallax should generally be utilized. This allows the viewer to perceive elements of the image both in front of and behind the display plane. The degrees of parallax, however, should be kept within certain limits. This paper describes methods for calculating both the initial settings and ongoing adjustments of stereoscopic parameters. Software using this approach should yield stereoscopic scenes that are comfortable and pleasing to look at, with a minimum of user adjustment.

We present a report on the effects of compressing stereo pairs using lossy compression techniques. We discuss the general functionality of all lossy compressors with particular emphasis on the CCITT/ISO JPEG Still Picture Compression Standard. We show that the application of lossy compressors to stereo pairs can cause ambiguities in the resulting stereo image and we develop a measure to quantify those ambiguities. We describe two new techniques which each help to decrease some of these anomalies. We also discuss the compression rates on some test images with and without the use of these new techniques.

A new method is proposed for extraction of 3-D rigid depth interpretations from pairwise comparisons of weak perspective projections. The method provides a simple criterion to test for the correctness of correspondence for a pair of images; the method also provides a description of a one-parameter family of interpretations for each pair of images that satisfies this criterion. We show that if at least three projections of a volumetric object are known, then a 3-D rigid interpretation can be inferred from pairwise comparisons between any one of these images and other images in the set. The 3-D interpretation is derived from the intersection of corresponding one-parameter families. The method provides a uniform computational basis for different processes of depth perception (for example, depth-from-stereo and depth-from- motion). In fact, a single mechanism for these processes in the human visual system could be sufficient. The proposed method does not require information about relative positions of eye(s) or camera(s) for different projections, but this information can be easily incorporated. This method can be applied for pairwise comparison within a single image. If any non-trivial correspondence is found it means that several views of the same object are present in the same image or that the object with volumetric symmetry is presented within the image. Results of pairwise comparison within a single image may also be considered in the process of depth reconstruction. If the object possesses two or more symmetries, its depth can be reconstructed from single image. Symmetry as a source of structural information is widely used by the visual system.

We identify two forms of error in stereoscopic vision: imprecision, which results from digitization of the images and imperfect knowledge of the system geometry and location in space; and uncertainty, generated by random sensor noise and the correspondence problem. Traditional approaches have either ignored one of the two sources of error or merged them into a single representation, namely a point probability distribution in space. Furthermore, unrealistic assumptions have been made on this distribution, which are constantly violated in practice. We propose a new mathematical framework, based on the random closed set theory, that allows for an explicit representation of both sources of error. Digitization and parameter imprecision produce an imprecision set in space, whereas noise and potential matching errors make the location of this set uncertain, hence probabilistic. We provide a practical method for estimating the probability distribution of this random set and demonstrate the efficiency of this approach on 3-D scene reconstruction from multiple stereoscopic views: errors in the final 3-D map are guaranteed to decrease in a statistical sense as additional views are processed. We introduce a connection machine parallel implementation where the 3-D space is represented in an octree. Experimental results on real images are presented.

A stereo display of a computer model adds relative depth information through binocular disparity. When the stereo viewing transformations are generated using knowledge of the viewer's head position, absolute depth positioning can be determined and a much more tangible presentation is created. The amount of benefit derived from such a display is dependent on the stability of the model's image as the head moves. To create this apparent stability the computer must display a different stereo view for each different position of the head, and update as often as possible. the image's stability is a direct result of both update rate and the ability to accurately determine the position of the head at the moment the stereo pair is presented. We present here the stereo transformations that we use and the results of our experiments in the area of accurately tracking the position of the head. We have tried various techniques to reduce noise in the stream of head position and orientation data. We have also tried prediction techniques on the same stream of data. We also describe applications of the technique, evaluating its utility with respect to different data types, data representations, and tasks.

Three-dimensional trackers are becoming increasingly important as user inputs in interactive computer systems. These trackers output the three-dimensional position, and often the orientation, of a sensor in space. The three-dimensional tracking is often, however, highly distorted and inaccurate. The purpose of this paper is to discuss methods for the measurement and characterization of the static distortion of the position data. When the distortion is constant, various methods can be used to calibrate the data from the tracker to increase accuracy. Several preliminary methods are discussed in this paper, including polynomial and weighted lookup methods. The measurement and calibration methods are applied to the Polhemus electromagnetic tracking system but are applicable to tracking systems based on other technologies.

Anti-aliasing is typically necessary for the satisfactory display of binary imagery on color matrix displays (CMDs) in order to attenuate visible spatial aliasing which can degrade image quality. One useful anti-aliasing approach is the use of gray scale (luminance quantization) to apply a discrete approximation of a Gaussian point spread function to CMD images. By convolving a Gaussian blur function with an image, prior to the sampling of that image, perceptible spatial artifacts such as stairstepping of lines can be significantly reduced. This paper examines the unique forms of spatial aliasing introduced in stereoscopic imagery and their impact on chromatic and spatial image quality. One form of stereoscopic aliasing, depth aliasing, varies as a function of the discrete sampling limitations of the 2-D horizontal axis of the display. Another form of stereoscopic aliasing, noncorrespondence aliasing, is governed by the noncorrespondence of 2-D spatial sampling artifacts between left and right stereoscopic images and may be exacerbated by chromatic disparities as well. The results of subjective image quality evaluations of simulated 2-D and stereoscopic flat panel displays are presented here with recommendations for the use of gray-scale anti-aliasing for the display of binary stereoscopic graphic imagery.

A series of three experiments examined temporal aliasing in stereoscopic displays. The first experiment compared aliasing in frontal plane motion of different disparities, while the second compared aliasing for motion in different depth directions. The results showed little effect of viewing conditions on perceived aliasing. The third experiment tested whether there was a binocular motion mechanism which integrated temporal sampling in the two eyes. The results were consistent with the first two studies in suggesting that aliasing is generated only by monocular motion signals. The data have both practical and theoretical implications: (1) motion produced by means of LCD glasses requires double the sampling rate needed for motion created by anaglyph methods, and (2) the short-range motion system is monocular.

Stereoscopic and monocular alignment acuity were measured using sinusoidal displacements in time, space, and disparity of a single line stimulus. The stereoscopic detectability was not limited by the sensitivity for the monocular components of the spatio-temporal stereo- alignment target. In fact, the two tasks seemed to be controlled by largely independent processes. Monocular sensitivity was best at high spatial perturbation frequencies, almost independent of temporal frequency, while stereoscopic sensitivity was best at low temporal and medium spatial frequencies, and its surface had a substantially different morphology. Under these dynamic conditions the lowest thresholds of either kind were of the order of 10 arc sec, setting stringent limitations on the accuracy of stereoscopic displays. The spatio- temporal surfaces we measured show regions where sensitivity is reduced by an order of magnitude, suggesting modes in which dynamic human stereopsis is more tolerant of perturbations than suggested by classical data.

The ability to make on-line adjustments to stereoscopic camera position parameters dynamically, during execution of telemanipulation tasks, allows one to maintain a theoretically `optimal' camera configuration, in response to changing viewing conditions. Associated with this, however, is the problem of the observer's being forced to adapt to a (continuously) changing relationship between perceived inter-object distances in the depth plane and the corresponding real distances. One problem in particular is the potential conflict between varying stereoscopic depth cues and unchanging size cues. Two experiments were performed. In the first we investigated how depth judgement ability varied with unsignalled changes in camera convergence distance. This resulted in significant changes in distance judgement, with overestimation for increases in camera separation and underestimation for decreases. Short- term feedback on judgement error was sufficient to correct the changes. In the second experiment, on-screen calibrated depth cues were added, by means of overlaid stereoscopic computer graphics, causing the significant estimation errors found in the first experiment to disappear. The implication of this is that distance/depth judgement can in principle be rescaled to compensate for perceptual conflicts caused by changing camera configuration, by providing either real or virtual depth scaling cues at the task site.

The efficacy of using point source lights to measure depth perception under remote view is evaluated. A Howard-Dolman type apparatus in which the depth plane is represented by either traditional rods or by point source lights is used. Ten operators, half viewing rods and half viewing lights, were asked to give depth scaling and stereoacuity judgments under four display conditions: (1) 2-D static, (2) 2-D motion parallax, (3) 3-D static, (4) 3-D motion parallax. The pattern of both stereoacuity and depth scaling responses was similar for rods and lights across the four conditions. Stereoacuity was significantly improved under 3-D as compared to 2-D view and under motion parallax as compared to static view for both lights and rods. When viewing lights but not rods a combination of motion parallax with disparity cues produced further improvements in stereoacuity. This pattern of results was similar for depth scaling, but differences were not significant. The accuracy of both stereoacuity and depth scaling judgments decreased when point source lights, as compared to rods, were viewed. These results show that point source lights produce valid measures of depth perception and contain fewer non-disparity cues than traditional Howard-Dolman rods.

Recently there has been significant activity in the attempt to develop autostereoscopic electronic displays. An interesting variation of the panoramagram, the moving slit technique, was described by Collender in the early seventies, and there have been various new types of volumetric display techniques, such as the Spacegraph acoustical mirror and the Texas Instruments laser scanned revolving surface. Lately liquid crystal technology has been employed by NTT and Dimension Technologies, offering the promise of a true three- dimensional display without the need for individual viewing devices. There are fundamental considerations with regard to presentation of visual information that provide constraints with regard to making such products competitive compared with current field-sequential electronic displays. These field-sequential displays have been successful in the marketplace and provide a standard against which the performance of new products must be measured. Products like CrystalEyesR allow any number of spectators to view the image, and have a high degree of compatibility with the present computer graphics and video infrastructures -- an important issue for manufacturers integrating such products into, for example, workstations, and for the user in terms of price and ease of use.

Dimension Technologies is currently one of three companies offering autostereoscopic displays for sale and one of several which are actively pursuing advances to the technology. We have devised a new autostereoscopic imaging technique which possesses several advantages over previously explored methods. We are currently manufacturing autostereoscopic displays based on this technology, as well as vigorously pursuing research and development toward more advanced displays. During the past year, DTI has made major strides in advancing its LCD based autostereoscopic display technology. DTI has developed a color product -- a stand alone 640 X 480 flat panel LCD based 3-D display capable of accepting input from IBM PC and Apple MAC computers or TV cameras, and capable of changing from 3-D mode to 2-D mode with the flip of a switch. DTI is working on development of a prototype second generation color product that will provide autostereoscopic 3-D while allowing each eye to see the full resolution of the liquid crystal display. And development is also underway on a proof-of-concept display which produces hologram-like look-around images visible from a wide viewing angle, again while allowing the observer to see the full resolution of the display from all locations. Development of a high resolution prototype display of this type has begun.

This paper reports on a 50 inch diagonal autostereoscopic full-color 3-D TV display system that requires no special glasses and uses four color TV cameras with a 3-D color viewfinder, a high-resolution liquid crystal (LCD) video projector, and a specially designed lenticular screen. The resulting 3-D image is of high resolution, exceptionally bright and vividly three- dimensional. The key technologies supporting this 3-D TV display system are a 3-D image display method, a newly developed HDTV LCD color video projector with incredibly high resolution (4.5 million pixels), and a new type of lenticular screen with a `hollow structure.' Images from four color TV cameras are recorded by four professional VTRs with a time code signal, and are multiplexed, pixel by pixel, to form a vertical stripe image projected by the single LCD video projector behind the lenticular screen. Using this 3-D image display method, optical alignment between the projected stripe image and lenticular screen is both accurate and easily attained. The new lenticular screen with hollow structure has resulted in a reduced total screen weight and thickness with increased light transmittance.

This paper describes a range of potential telecommunications applications for 3-D imaging systems. We outline possible techniques to realize 3-D video and still image demonstrations using autostereoscopic systems. These are glasses free systems involving the use of view determining screens, either parallax barriers or lenticular sheets. Results showing the optical performance of lenticular sheets and parallax barriers with test still images are described and analyzed. Good agreement with theoretically derived models has been obtained.

In designing remote vision systems for human operators, two different features of the human eye should be considered: (1) the very wide field of view of the eye permits secure and rapid orientation and the avoiding of obstacles; (2) the high central (foveal) acuity of the eye permits fine discrimination. A capable head-mounted display will not limit either of these human eye features more than absolutely necessary. LCDs show promise for reasonable cost and weight, but the small number of pixels available now and in the foreseeable future means a necessary limitation of the field of view or of the central acuity, or of both, if only one LCD is used per eye. The best available compromise using only one LCD is embodied in the LEEP optics, which use a radial compression format to provide a very wide field with minimum loss of resolution at the center. LEEP Systems, Inc. has undertaken two developments. The first simply increases the lateral field of view by twenty-five degrees (the CYBERFACE2). The second adds an optical high-resolution insert. Because the insert is added optically in the head- mount, the smaller the insert, the sharper its detail.

That stereopsis is scalable is often misunderstood. Because stereopsis is scalable, the Carter single lens system offers many advantages over dual lens imaging in the accuracy and timing of mechanical tasks. Before a typical application is described, the bias against, and the human factors issues surrounding stereopsis are examined. This single lens solution is offered and its application to microscopic probes is detailed.

Existing high-resolution monitors are optimized for display of non-stereoscopic images with field refresh rates of 60 to 80 hertz. Almost all existing graphics systems utilize refresh rates in this range. Stereoscopic field-sequential displays present alternate left and right images, with each eye seeing half the displayed fields by use of electronic shuttering systems. This image selection is accomplished by optical shutters that are alternately clear and opaque operating synchronously with the display. To maintain flicker-free display for each eye requires at least the doubling of the existing field rate. An idealized monitor for stereoscopic display adds several new demands on the performance of monitors that extend beyond existing requirements. Some of the new requirements may be contrary to existing needs, calling for engineering compromises to be considered. The paper addresses the electronic and perceptual requirements of stereoscopic monitors in the areas of scan ranges, phosphors, and interfaces. Success in utilizing existing commercial monitors and projectors and possible future directions are discussed.

Following some comments on the nature of stereo perception as it relates to stereo video displays, a number of areas of interest are briefly reviewed and accompanied by extensive citations from the patent and technical literature. These include various methods of autostereoscopic display, stereoendoscopy, stereo sculpture, and holographic television.

A low cost helmet-mounted stereoscopic color viewing system designed for field testing teleoperator tasks is described. A stereo camera pair was mounted on a helmet to allow testing of a helmet-mounted display with real time video input. The display consisted of a pair of LCD color monitors viewed through a modified Wheatstone mirror system. The components were arranged on a stable platform that was attached to a hard plastic helmet. The helmet weight (9.5 pounds) was supported by a modified backpack. This backpack also contained support electronics and batteries. Design, construction, and evaluation tests of this viewing system are discussed.

Space Shuttle missions and future Space Station Freedom activities both require a significant amount of out-the-window viewing to perform a variety of on-orbit tasks such as grappling objects with a robot arm, berthing payloads into the shuttle cargo bay, and docking the shuttle to the station. Many of these tasks take place close enough to the viewer such that stereo depth cues become useful and important. Astronaut crews spend many hours of ground training practicing these tasks using visual simulation without stereo depth cues. A prototype of a stereoscopic display system for use in these training scenarios has been constructed. Training instructors and development contractor personnel were asked to evaluate the display system for training suitability. Evaluators were also asked to judge the quality of stereo cues, long term viewing comfort, and degree of display artifacts. Results were compared to related research to develop recommendations for scene content and viewing distance and suggest the direction of further research to improve comfort and depth perception.

We discuss several representation techniques of motion pictures on television with respect to a possible repression of parallactic depth cues and an accentuation of the monocular cues. Different viewing aids which promote this intention are presented.

One of the objectives of this study was to determine which of two cursor control devices would provide the best means of designating targets in stereo three-dimensional (3-D) space. The two devices tested were a joystick and a hand tracker. The second objective was to determine if performance improved with the use of an aiding technique to designate the targets. This aid was incorporated by changing the color of the cursor when it penetrated the target volume. The total 3-D viewing volume was divided into four perceived depth volumes within which the targets could appear. The study showed that the quickest designations occurred when using the hand tracker, and the hand tracker performed significantly better than the joystick when the targets were located in the two volumes behind the screen. There was no difference in accuracy for the two devices. Also, the use of the aiding technique was most effective (target designation time and accuracy) in the farthest behind depth volume and the front-most depth volume. Performance in the other depth volumes showed no advantage of aiding.

Computer animation consists of a series of images each taken for a specific point in time. This can create temporal aliasing; motion, especially fast motion, appears discontinuous. Motion blur can be used to prevent a strobing effect and to enhance the perception of motion. Motion blur is commonly achieved by one of the following three methods: stretching the object along the path of motion, stochastic sampling in the time domain, and supersampling over time. When creating stereo images while using stochastic sampling, each eye sees different points on the object at a pixel location over time, creating discrepancies in the left and right views. Supersampling produces images that are averages of several disparate images of the same object, making it difficult to fuse the views. Traditional methods for creating motion blur can therefore produce images with ambiguous depth when combined with stereo. The stretched object method works well with stereo as there is consistency in both views. However, it exaggerates sizes of moving objects and lacks the blurring effect of the sampling methods. Our technique for creating motion blur in stereo is to stretch moving objects and then apply functions to give blurring effects such as transparency or fading.

Recent developments in virtual 3D audio and synthetic aural environments have produced a complex acoustical room simulation. The acoustical simulation models a room with walls, ceiling, and floor of selected sound reflecting/absorbing characteristics and unlimited independent localizable sound sources. This non-visual acoustic simulation, implemented with 4 audio ConvolvotronsTM by Crystal River Engineering and coupled to the listener with a Poihemus IsotrakTM, tracking the listener's head position and orientation, and stereo headphones returning binaural sound, is quite compelling to most listeners with eyes closed. This immersive effect should be reinforced when properly integrated into a full, multi-sensory virtual environment presentation. This paper discusses the design of an interactive, visual virtual environment, complementing the acoustic model and specified to: 1) allow the listener to freely move about the space, a room of manipulable size, shape, and audio character, while interactively relocating the sound sources; 2) reinforce the listener's feeling of telepresence into the acoustical environment with visual and proprioceptive sensations; 3) enhance the audio with the graphic and interactive components, rather than overwhelm or reduce it; and 4) serve as a research testbed and technology transfer demonstration. The hardware/software design of two demonstration systems, one installed and one portable, are discussed through the development of four iterative configurations. The installed system implements a head-coupled, wide-angle, stereo-optic tracker/viewer and multi-computer simulation control. The portable demonstration system implements a head-mounted wide-angle, stereo-optic display, separate head and pointer electro-magnetic position trackers, a heterogeneous parallel graphics processing system, and object oriented C++ program code.