Paul Milgram & Martin Krüger
Department of Industrial Engineering
University of Toronto, Toronto, Canada M5S 1A4
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 research presented here was motivated by the potential for making on-line adjustments to stereoscopic camera parameters dynamically, that is, during execution of a telemanipulation task. Most research to date on related topics has concentrated on investigating the advantages of using stereoscopic video systems for teleoperation, in comparison with monoscopic video or multiple camera systems (e.g. Cole and Parker, 1989; Merritt, 1984), and on how various camera and task parameters are related (e.g. Spain, 1984). In his study of skill acquisition and task performance using stereoscopic viewing systems, for example, Drascic (1991) emphasised the important factor of the stereo dependency of the particular task site being viewed. Notwithstanding a number of inconsistencies arising from early studies in this area, it is generally well accepted that using stereoscopic displays provides a number of potential benefits, including faster and more accurate perception of the spatial layout of the remote scene, visual noise filtering, enhanced effective image quality, and enhanced object recognition and image interpretation (Merritt, 1988).
Nevertheless, the "solution" to the practical problem of how precisely to specify a suitable camera separation, convergence angle, focal length, elevation, viewing angle, etc. for any particular stereo dependent task remains elusive. Results of Spain's (1984) studies showed that there is no single "optimal" camera setup. This has prompted a number of proposals for the fabrication of adjustable camera systems (e.g. Scheiwiller et al, 1991; Drumbeck et al, 1988; Goode & Philips, 1989), which should, in principle, allow the user to adapt camera parameters to the particular viewing circumstances. Surprisingly, however, there have been few studies published thus far in which performance results obtained using such adaptable systems have been reported.
The research presented here is based upon the premise that an adjustable-base converged camera stereoscopic viewing system can be useful, as well as technologically feasible, but that a database of empirical results is necessary to enable prescriptive design of such systems. In this paper we address the effects on human perception of changing one of the basic stereo parameters: the separation between the two cameras. By adjusting this parameter as required, an operator should in principle be able:
to decrease the camera separation at some times, thereby accepting a low level of depth resolution, but gaining in terms of ability to fuse and perceive a large range of objects in depth, and
to increase the separation at other times, thereby gaining an increase in depth resolution, but losing in terms of range of fusable images.
Fig. 1 shows the general setup of a converged camera stereoscopic viewing system. Essentially there are four independent parameters of such a system which will affect the resulting stereoscopic image on the CRT monitor. Two of these are shown in Fig. 1: the intercamera separation, Ic, and the convergence distance, Dc. While the convergence distance, Dc, is often functionally the one that is used for adjustment, this parameter is actually a combination of the camera convergence angle, ac, and camera separation, Ic. The other two parameters (not shown) are the focal lengths, f, of the cameras and the CRT monitor magnification factor, M, both of which have related effects.
Figure 1. Stereoscopic camera system and design parameters
Table 1 presents a summary of the effects of varying the first three of these parameters individually. (Further details on these factors, as well as a discussion of the formulae that relate them, can be found in (Krüger, 1991).) The two parameters Ic and f both allow one to increase the depth resolution of the displayed image (i.e. increase of disparity values), but a change of camera separation (Ic) will preserve the field of view, whereas a change of focal length will not. The problem associated with not having an accompanying change in the field of view in the former case is that objects will not appear to become bigger as camera separation is increased, and a cue conflict will result between size and disparity information. That is, the extended depth percept will suggest that the distances between objects in the depth direction has increased, whereas the constant size information will imply that these distances have remained constant. This is not the case with changes in focal length (f).
Table 1: Summary of effects of camera design parameters (from Krüger, 1991)
In spite of the abovementioned cue conflict, adjusting the camera separation (Ic) can still be useful for changing the resolution of a displayed image, that is, for causing a particular physical distance to appear to be larger whenever Ic is increased. The disadvantage of using focal length (f) for this purpose is that, because field of view decreases as focal length increases, it is not always possible to maintain extended depth resolution and still observe all objects in the image. In addition, changing the focal lengths of both cameras continuously and simultaneously is technologically very difficult (Scheiwiller et al, 1991).
Increasing the convergence distance (Dc), by rotating the two cameras away from each other, will have the effect of causing all objects which are in front of the plane of convergence (crossed polarity) to move forward out of the screen. As with camera separation, however, this shifting is not accompanied by a concomitant change in size, so that a similar cue conflict will result. Nevertheless, adjusting the convergence distance can be a useful tool if it is required to put the target of a particular task at the plane of the screen, in order to avoid mismatches between accommodation and convergence (although this strategy is not always supported by performance results (Spain, 1984)) and to minimise stereoscopic depth distortions (Diner & von Sydow, 1987).
Very little experimental research has been carried out on the effects of on-line adjustment of these camera parameters on actual task performance, a reflection perhaps of the paucity of practical human factors research in teleoperation viewing systems in general, as well as of the small number of existing adjustable stereoscopic camera systems in particular. Of the few such systems that have been described in the recent literature (e.g. Drumbeck et al, 1988; Scheiwiller et al, 1991; Goode & Philips, 1989; Robinson & Shuttleworth, 1988) which are capable of adjusting at least two of the three camera parameters, no publications were found on how performance is or will be affected by such on-line adjustments. Research at the Naval Ocean Systems Centre (Spain, 1984; Pepper and Hightower, 1984; Pepper et al, 1986) on discretely varying systems has shown that increased levels of camera separation and magnification can be usefully employed, but the extent depends very much on the actual task at hand. That is, a large camera separation will have advantages for some tasks, such as matching objects in depth, while it will be less useful for other tasks, such as detecting camouflaged objects, as compared with magnification. No studies were found, however, which have investigated how performance is influenced by changing the stereoscopic camera setup during a task, rather than between tasks.
In addressing this problem, we have adopted the paradigm of distance/depth scaling, which has been used in a number of basic investigations (e.g. Fisher & Ebenholtz, 1986; O'Leary & Wallach, 1980) as well as applied studies (e.g. Spain, 1984; Gaillard, 1988) on depth perception. Two experiments were carried out, with the following three general objectives:
1. to identify the effects of changes in disparity-depth relationship, i.e. a changing camera separation, during a depth scaling task;
2. to investigate whether feedback for a short period of time will result in adaptation, i.e. a reduction of any adverse effects of unfamiliar or newly changed camera separation;
3. to investigate whether the addition of calibrated depth cues to the scaling task can counter any adverse effects of changed camera separation.
From the second and third of these objectives it is clear that our hypothesis regarding the first objective was that performance would indeed deteriorate whenever camera separation is changed during a depth scaling task, especially if this occurs unbeknownst to the subject. In other words, we predicted that the conflict between binocular disparity and size constancy cues would in fact affect performance. The motivation behind the last two questions, therefore, was to make an initial attempt at coming up with a practical solution to these problems.
In each experiment, subjects were given a depth scaling task, for which they were required to judge the distance between each of a series of target objects and a reference object, viewed on a stereoscopic video system. Figure 2 shows a frontal view of (a monoscopic representation of) a sample screen with these stimuli. Each object was two dimensional, made from white cardboard, with dimensions 10 cm x 10 cm. Their lateral separation was also 10 cm. The left square always had a red symbol "L" on it, and the right square an inverted L => "G". The actual separations between the targets and the central reference object, i.e. the cross shown in Figure 2, ranged over trials between -22 and +22 cm. Subjects were required to estimate the distances in absolute units, i.e. cm.
Figure 2: Example of stimulus screens for Experiment 1
Because there were no cues available for the subjects to base such absolute judgements upon that is, simply viewing two arbitrary squares, which are clearly separated in depth, is not sufficient information for judging how far apart they are feedback was provided over a portion of the trials. The experimental procedure is summarised in Figure 3. Full details of the experiment are given in Krüger (1991).
Figure 3. Stimulus ordering procedure for Experiment 1
At the top of Figure 3 is indicated that 56 separate data screens were presented to each subject. Each experiment began with a Low camera separation. For each of the first 16 trials, subjects were shown a screen and were first asked to estimate the separation in depth between the Left square and the central reference cross, as illustrated in the left part of Figure 2. Because the objects displayed appeared essentially as two dimensional silhouettes displaced in depth, they did not provide any absolute depth information on their own. Consequently, after each response was entered, an additional text display appeared indicating how close the estimate was to the real distance, as illustrated in the right part of Figure 2. This type of presentation comprised the first "Feedback" segment of the experiment, indicated in Figure 3. Five seconds after the feedback was given, the subject was required to make a similar judgement for the distance between the Right square and the reference cross.
After the 16th data screen, the experiment continued as above, except that no feedback was provided for the following 6 data screens. This comprised the first "Post Feedback" segment of the experiment. The purpose of this segment was to allow the subject to achieve an acceptable level of performance, by providing feedback, and then to evaluate the level of performance after removal of the feedback.
Following the 22nd data screen, the camera separation was increased between trials, to either a Medium or a High separation, without the subject being told, and the experiment continued as before. The following 6 data trials (#23-28) were similar to the preceding 6 screens (#17-22), in that no feedback was provided, in order to permit assessment of the impact of the change in camera separation. This segment is labelled "Pre Feedback" in Figure 3.
Following the Pre Feedback phase under increased camera separation, subjects underwent another 16 trials (#29-44) with Feedback, to allow them to reach an acceptable level of performance with the new camera separation. After this, as illustrated in Figure 3, another Post Feedback (i.e. no feedback) phase was administered. This was followed by another unsignalled change in camera separation, this time a decrease to the original Low separation, and a final Pre Feedback (i.e. no feedback) phase for comparison purposes. In total, 112 (2x56) judgements were made by each subject.
The bottom row of Figure 3 indicates the important transitions that were of interest for the analysis of the results. In accordance with Objective 1 listed earlier, changes in performance over Transitions A and B were used as indicators of the effects of the changes in camera separation: Transition A for the increase and Transition B for the decrease. Analysis of Transition C is designed to address objective 2, that is, whether providing feedback for a short period of time is sufficient to overcome the effect of the changing camera parameters. (Objective 3 is addressed in Experiment 2.)
In light of our research objectives discussed earlier, the following hypotheses were adopted for Experiment 1. (The letters A, B, C correspond to the transitions shown in Figure 3.)
1A Switching from a smaller to a larger camera separation that is, increasing depth resolution will result in overestimation of absolute depth.
1B Switching from a larger to a smaller camera separation that is, decreasing depth resolution will result in underestimation of absolute depth.
2 Provision of (short-term) feedback of depth scaling error will be sufficient for adaptation in depth judgement to occur that is, for depth scaling errors to be effectively reduced.
A field sequential stereoscopic camera system, consisting of two Hitachi VK-C150 colour CCD cameras, equipped with 16 mm Cosmicar lenses, was used. The cameras were positioned on an adjustable camera mount at the far end of an enclosed darkened booth. The camera mount enabled continuous adjustment of the camera separation and convergence distances by means of two stepping motors, which in turn were controlled by a computer program via a stepping motor driving circuit. Three different camera separations were used:
Low = 80 mm
Medium = 160 mm
High = 240 mm
For each separation the camera convergence distance was set to 1.782 m, which coincided with the position of the central reference object (the cross), shown in Figure 2.
An Amiga 2500 computer was used to control the experiment, including driving the stepping motors and displaying the stereoscopic images. The latter was accomplished by genlocking the NTSC video signal from the cameras, using a Commodore 2300 genlock unit, and alternating odd and even fields on an interlaced monitor, a 240 x 270 mm Commodore 1084 colour monitor. Subjects wore Haitex liquid crystal shuttering spectacles to perceive the images in stereoscopic depth. A chin rest was used to constrain head position, and all images were viewed from a fixed viewing distance of 0.85 m. At this distance the visual angle of the square objects was approximately 2°x2°. All subject responses were given by means of a MicroSpeed FastTrap Trackball + Thumbwheel.
Twelve students (7 male and 5 female), between the age of 20 and 30, served as subjects in the experiment. They were paid $25 for participating in Experiment 1 and 2, or $7 for Experiment 1 only. None had any experience in using stereoscopic displays. All subjects were required to pass a stereoscopic vision screening test, based on depth ranking of objects and fusion of images. Depending on their ease of fusing large disparities, subjects were divided into two groups, which accordingly received either the Medium (Experiment 1a) or the High (Experiment 1b) camera separation during the experiment.
All response data were transformed into distance judgement error values, by subtracting the subject responses from the actual object separations: Judgement Error = Subject Estimate - Actual Separation.
These data were analysed using 3- and 4-way repeated measure ANOVAs, provided by the Powerstat analysis package. Only a portion of the analysis is presented here; details of the full analysis can be found in Krüger (1991). It is important to point out that a randomised design was not used, and thus results are potentially confounded by sequence effects (see Section 5).
One main effect during the Feedback phase that was significant in both parts of the experiment was that of the distances separating the targets. (In Expt. 1a: F(7,35)=9.62, p<.0005 and in Expt. 1b: F(7,35)=14.882, p<.0005 ). As illustrated in Figure 4, smaller target separations result in overestimation (i.e. positive judgement errors) and larger separations cause underestimation (i.e. negative judgement errors). The effect of resolution (i.e. either Medium or High camera separation) was only marginally significant.
Figure 4: Judgement errors during Feedback phases of Experiments 1a and 1b .
Using a 3-way repeated measures ANOVA (Camera Separation x Pre/Post-Feedback x Target Distance) to examine the main transitions of interest, it was found that, for both parts of the experiment, both Camera Separation ( F(1,5)=8.93, p<.03 for Experiment 1a and F(1,5)=80.6, p<.0005 for Experiment 1b ) and Target Distance ( F(5,25)=4.47, p<.005 and F(5,25)=13.7, p<.0005 respectively) had significant main effects. Similar to the results shown in Figure 4, the effect of distance is again an overestimation of small distances and underestimation of larger distances.
A number of two and three way interactions were examined (see details in Krüger, 1991). Among the most interesting is the interaction between Camera Separation and Pre/PostFeedback, for both Experiments 1a and 1b, summarised in Figure 5. (For Experiment 1a, F(1,5)=3.62, p<.115; for Experiment 1b, F(1,5)=27.5, p<.003.) Figure 5 shows the effect of how judgement errors were affected by each change in Camera Separation (Transitions A and B) and by the administration of feedback (Transition C), for both sets of increased camera separations (Expt. 1a => Medium and Expt. 1b => High). All four sets of Post Feedback results are very small, indicating that the administration of feedback was indeed effective. (This is further discussed below: Fig. 6.)
Figure 5: Judgement errors resulting from Camera Separation x Pre/Post-Feedback.
We now examine the effect of changing the camera separation, corresponding to the transitions from Post-Feedback to Pre-Feedback i.e. Transitions A and B. First of all, we note the dramatic increase in judgement error, to positive errors, for Transition A, corresponding to overestimation in distance judgement. These errors are in general smaller for the Low>Medium transition (Expt. 1a) than for the Low>High transition (Expt. 1b). Examining the effect of Transition B, we see that the very small errors in the Post-Feedback phase, which resulted from administration of the feedback, disappear after the cameras return to their original Low separation. This time, however, we observe underestimation, that is, generally negative judgement errors, which are once again larger for the Medium>Low transition than for the High>Low transition. A posthoc analysis, using a Student-Newman-Keuls Multiple Range test, verified that all Pre-Feedback values are significantly different from the Post-Feedback values (p<.05). These results thus support Hypotheses 1A and 1B, presented in Section 3.2.
To examine further the issue of the effectiveness of error feedback for adaptation in depth scaling, which is suggested in Figure 5, we present Figure 6, which shows the depth judgement errors during the Post- and Pre-Feedback phases, for Experiment 1b. The left hand side of this figure, indicated by the negative trial numbers, shows the low judgement errors prior to the change in camera separation. The vertical dotted line indicates the change in camera separation, followed by 6 subsequent screen judgements (positive trial numbers). For Transition A (labelled "Low -> High-Pre"), judgement errors are consistently positive, as discussed above. For Transition B (labelled "High -> Low-Pre"), judgement errors are consistently negative. The message to be derived from this graph is that without feedback no significant improvement in judgement ability occurs. In fact, when subjects were asked after the experiment about the change in camera separation, almost all of them appeared not to have been aware at all that any change had taken place. These results, along with those presented in conjunction with Figure 5, are therefore taken to support Hypothesis 2, presented in Section 3.2.
Figure 6: Pre-Feedback phase judgement errors, trial-by-trial, Experiment 1b.
4.1 Objectives and Hypotheses
Reiterating objective 3, presented in Section 2 above, the aim of this part of the research was to investigate whether the addition of calibrated depth cues to the scaling task could be useful in countering any adverse effects of changed camera separation. Recall that in Experiment 1 the subjects observed the silhouettes of two 2D squares displaced in depth, but had essentially no means for judging the absolute distance between them other than by means of the error feedback training provided. In Experiment 2, on the other hand, subjects were presented with one of two different kinds of reference objects, which provided a known standard of comparison in the depth plane. In addition, subjects were presented an explicit indication of the magnitude of the camera separation at all times. Both of these additions were intended to simulate a more "realistic" setting.
An example of each of the visual aids is shown in the monoscopic representation in Figure 7. On the left hand side the visual aid, shown in perspective depth, is a 30 cm long elongated box, with a 4 x 4 cm cross section, symmetrically placed in depth with respect to the reference cross. On the right hand side the corresponding visual aid comprises two 4 x 4 cm boxes, each placed 15 cm in front of and behind the reference cross. Note also in Figure 7 that subjects are explicitly informed about camera separation (Depth Resolution), as being either LOW or HIGH. Note, finally, that the feedback in this experiment was presented in a "fuzzy" form, rather than as exact numbers, in order to inhibit subjects' ability to memorise the set of displacement stimuli.
In light of the discussion above, the following hypotheses were adopted for Experiment 2:
3A/B Switching between different camera separations (Transitions A and B) will not result in respective over- or under-estimation of displacement, due to the presence of the strong depth/distance cues provided.
3C Error corrective feedback will not result in adaptation (Transition C), which will already have occurred in each Pre-Feedback phase due to the depth/distance cues provided.
Figure 7: Example of stimulus screens for Experiment 2
4.2 Method, Apparatus, Subjects
The experimental task in Experiment 2 was identical to that of Experiment 1. Ten subjects (5 male and 5 female) between the age of 20 and 30 took part. All of them had participated in Experiment 1, at least 2 days earlier. The apparatus was identical. Only the 160 mm camera separation ("Medium" in Expt. 1) was used, due to the difficulty that some subjects had in fusing images with the High separation in Experiment 1. Camera separation was indicated explicitly and stimuli were as illustrated in Figure 7, with one of the two aids presented. The form of the feedback was "fuzzy", as mentioned above. The design of the experiment was essentially the same as Experiment 1, with only minor differences that will not be discussed here.
The effect on judgement of the actual distances between the targets during the Feedback phase was similar to that of Experiment 1 (Fig. 4), with a significant main effect of distance (F(13,117)=7.66, p<.0005) and a significant interaction between distance and camera separation (F(13,117)=6.98, p<.0005). Although a similar conclusion was drawn, that large target separations tend to be underestimated and small ones overestimated, it is important to note that the magnitude of the errors in this case (see below) was much smaller than in Expt. 1.
Figure 8: Judgement errors resulting from Camera Separation x Pre/Post-Feedback.
A four-way repeated measures ANOVA for the Pre/Post Feedback segment of the experiment revealed a significant main effect of Camera Separation (F(1,9)=18.2, p<.002) and Distance (F(5,34)=4.64, p<.002), as well as a two-way interaction between Camera Separation and Pre/Post Feedback (F(1,9)=9.21, p<.014). The latter interaction is illustrated in Figure 8, the principal result of this experiment, which shows how judgement errors were affected by the three transitions of interest.
First of all, we note in Figure 8 that, in comparison to Figure 5 for example, the magnitude of errors is very much smaller over the whole range of conditions. This is attributed to the presence of the visual aids provided. Secondly, we note the results of a post-hoc Student-Newman-Keuls Multiple Range test, which showed that, even though the two Pre-Feedback measures are significantly different from each other (p<.05) and indicate a slight overestimation for the Medium camera separation and a slight underestimation for the Low separation, neither of them is significantly different from the Post-Feedback values. Together these results lend support to Hypotheses 3A/B and 3C presented above.
In the interest of brevity, we shall not repeat the various hypotheses listed above. As discussed in the preceding text, essentially all of them were supported by the results of the two experiments.
As stated earlier, the present report is a subset of the results obtained in a larger study (Krüger, 1991). Some of the other conclusions reached from Experiment 2 include the following:
Subjects preferred the large camera separation conditions to the low camera separations, even though there were no significant performance differences in the Post-Feedback phases.
Subjects preferred the elongated box to the separated markers, even though there were no performance differences.
Visual aids which were presented in Experiment 2 as overlaid stereo-graphic images were just as effective as visual aids which were implemented as real objects within the live stereo-video scene.
One practical aspect of this experiment which must be mentioned is the weakness of portions of the experimental design. In both experiments it was necessary to present each subject with the same stimuli in the same order. Consequently, it is conceivable that the results obtained could have been confounded with practice effects during the experiment. Further confounding could be argued to have occurred in Experiment 2, where not only were the subjects already experienced, after having participated in Experiment 1, but the addition of two new depth cues the explicit visual elongated box/separated marker aids plus the indication of camera separation could be used as an argument against identifying any one factor as the most effective factor. Being quite aware of these potential shortcomings, our motive in designing the experiments as described was first to explore in Experiment 1 the basic "problems" inherent in carrying out depth scaling in a depth cue impoverished environment and then to present a candidate approach in the more "realistic" setting of Experiment 2, which combined more than one means for abetting the operator in carrying out the task.
This brings us, in conclusion, to consider some of the real-world implications of this research. One implication is in support of the practicality of using an adjustable based stereoscopic camera system in real remote operations, to benefit from some of the advantages outlined earlier. Although in our experiments we did not actually employ dynamic changes in camera separation, we did nevertheless show that it is possible to adapt rather quickly to some changes in the camera setup and thus overcome some of the problems concomitant with such on-line changes. Furthermore, it was not surprising to find that, when provided with visual depth cues of known/given dimensions, subjects could use this information as an anchor for making reasonably accurate mental estimates of depth displacement. In a well-structured, familiar environment the controller of a teleoperated manipulator or vehicle would surely employ such cues. However, what sorts of options are available for operations in unfamiliar environments, or at remote or hazardous worksites at which few or no such cues are present?
On the basis of the results presented here, one novel option, which has only recently become technologically feasible, would be to superimpose onto the stereoscopic video image of the remote scene a set of calibrated markers, "rulers", or any other suitable depth cues required, by means of overlaid stereoscopic computer graphic images of those markers, rulers, etc. Further results and experiences with such display system features are reported in Milgram et al (1990; 1991) and Zhai & Milgram (1991).
This work was funded jointly by the Defence and Civil Institute of Environmental Medicine (DCIEM), Downsview, Ontario, Canada, under contract W7711-7-7009/01-SE, with Dr. J.J. Grodski as scientific monitor, and by the Deutscher Akademischer Austauschdienst (DAAD) in Bonn.
Cole, RE & Parker, DL (1989). "Stereo TV improves manipulator performance", in SPIE Vol. 1083: Three-Dimensional Visualization and Display Technologies, 18-27.
Diner, DB & von Sydow, M (1987). Stereo Depth Distortions in Teleoperation. NASA JPL Publication 87-1.
Drascic, D (1991). "Skill acquisition and task performance in teleoperation using monoscopic and stereoscopic video remote viewing", in Proc. Human Factors Society 35th Annual Meeting, San Francisco, 1367-1371.
Drumbeck, AA, Abel, E, Holt, J, Scheiwiller, PM & Reading, VM (1988). "A 3-D television system for remote handling", in Proc. of International Symposium on Teleoperation and Control, Bristol, UK, IFS Ltd., 197-207.
Fisher, SK & Ebenholtz, SM (1986). "Does perceptual adaptation to telestereoscopically enhanced depth depend on the recalibration of binocular disparity?", Perception & Psychophysics, 40(2), 101-109.
Gaillard, JP (1988). "Operator's distance estimation to supervise satellite grasping", in Training, Human Decision Making and Control, J Patrick & KD Duncan (eds), Elsevier, 383-393.
Goode, GD & Philips, ML (1989). "Voice controlled stereographic video-camera system", in SPIE Vol. 1083: Three-Dimensional Visualisation and Display Technologies, 35-41.
Krüger, M. (1991). Adaptation in Depth Perception using Stereoscopic TV Displays. M.A.Sc. Dissertation, Dept. of Industrial Engineering, University of Toronto.
Merritt, JO (1984). "Visual tasks requiring 3-D stereoscopic displays", SPIE Vol. 462, Optics in Entertainment II.
Merritt, JO (1988). "Often-overlooked advantages of 3-D displays", in SPIE Vol. 902, Three-Dimensional Imaging and Remote Sensing Imaging, 46-47.
Milgram, P, Drascic, D & Grodski, JJ (1991). "Enhancement of 3D video displays by means of superimposed stereo-graphics", Proc. 35th Annual Meeting of Human Factors Society, 1457-1461.
Milgram, P, Drascic, D & Grodski, JJ (1990). "A virtual stereographic pointer for a real three dimensional video world". In D Diaper et al (Ed's): Human-Computer Interaction INTERACT '90, Elsevier Science, 695-700.
O'Leary, A & Wallach, H (1980). "Familiar size and linear perspective as distance cues in stereoscopic depth constancy", Perception & Psychophysics, 27(3), 131-135.
Pepper, RL, Spain, EH & Cole, RE (1986). "Design issues in teleoperator vision systems", in SPIE Vol. 624, Advances in Display Technology VI, 36-40.
Pepper, RL & Hightower, JD (1984). "Research issues in teleoperator systems", in Proc. Human Factors Society 28th Annual Meeting, 803-807.
Robinson, M and Shuttleworth, P (1988). "The development of stereoscopic vision systems for use in hazardous environments", in Proc. of International Symposium for Teleoperation and Control, Bristol, UK, IFS Ltd.
Scheiwiller, PM, Murphy, SP & Dumbreck, AA (1991). "Compact zoom lens for stereoscopic television", in SPIE Vol. 1457: Stereoscopic Displays and Applications II, 2-8.
Spain, EH (1984). A Psychophysical Investigation of the Perception of Depth with Stereoscopic Television Displays; Ph.D. Dissertation, University of Hawaii.
Zhai, S & Milgram, P (1991). "A telerobotic virtual control system", in SPIE Vol. 1612, Intelligent Robotic Systems: Cooperative Intelligent Robotics in Space II.