3.1 Parallax distribution and visual comfort of stereoscopic images
One of the issues that can arise when making stereoscopic images widely available to large numbers of viewers is the visual discomfort experienced by viewers of some scenes. This discomfort is thought to be caused by discrepancies and crosstalk between the characteristics of the images presented to the left and right eyes, such as differences in their geometrical characteristics or video properties, and is thought to be more of a problem in scenes where these characteristics are prominent. Another factor is parallax itself, which plays a key role in conveying the sense of depth.
It has often been noted that excessive parallax can cause visual fatigue. There are many productions where surprisingly large amounts of parallax are used intentionally for dramatic impact. If we can find out about the distribution of parallax within the same frame and how this distribution affects the viewer's visual discomfort, then this should provide us with very useful clues for the production of visually comfortable stereoscopic images in each scene. It can sometimes be difficult to analyse the parallax values included in a frame. In what follows, we describe such an analysis and apply it to a number of stereoscopic images. We then compare its results with those of subjective evaluation tests.
3.1.2 Parallax measurements
Here we summarize the parallax measurement method used in this study.
For many applications, we ideally need to be able to determine the depth (i.e. the amount of parallax) for each pixel of an image. However, algorithms for analysing parallax are generally prone to errors. Also, in practice there are cases where corresponding pixels do not exist in the images presented to the left and right eyes due to occlusion. In this study, we do not need to measure parallax on a strict per-pixel basis as long as we are able to extract the characteristics of a parallax distribution from the images. To achieve this aim, the method we propose combines phase correlation with a number of threshold processing methods. A detailed description of this algorithm can be found in the Reference below. The algorithm was used in the parallax analysis of stereoscopic images discussed below.
3.1.3 Subjective evaluation tests of parallax distributions and visual comfort
We performed subjective evaluation tests to investigate the relationship between visual comfort of stereoscopic images and their parallax distributions . The subjective evaluation test conditions are shown in Table 7. The images used for the evaluation consisted of 48 different still images.
These images were presented as stereoscopic images with 2D images as a standard reference, and their relative visual comfort was evaluated on a seven-grade scale. The 2D images were produced by presenting the left-eye image to both eyes and were evaluated by test subjects wearing the same polarizing glasses used for the stereoscopic images. The parallax in the stereoscopic images was measured by using the phase correlation method discussed in § 2. Here, the amount of parallax was measured on a per-pixel basis with the screen corresponding to a value of zero, and positions behind and in front of the screen correspond to positive and negative parallax values respectively. In the viewing conditions shown in Table 7, the amount of parallax of a single pixel corresponds to a separation of approximately 1 mm between the left and right images on the screen.
Figure 11 shows the results of the visual comfort subjective evaluation tests and the results of measuring the amount of parallax in the images. In the graphs of this Figure, the numbers on the horizontal axis designate the images to be evaluated. The graph at the top shows the results of the subjective evaluation tests. The vertical axis shows the mean value of the evaluation scores from 24 evaluators, and the vertical bars represent the standard deviations of these scores. The lower graph shows the results of measuring the amount of parallax. The vertical axis shows the amount of parallax measured in pixel units, the plotted points show the average values of the parallax in the images, and the vertical bars show the range of the parallax distributions. The upper and lower ends of the vertical bars represent the maximum and minimum parallax values. A comparison of the two graphs shows that the images with an evaluation score of 3 or less have a very large parallax distribution range.
Conditions of subjective evaluation tests of parallax distributions
Images used in test
48 still images (including a standard pattern)
24 adult males and females (not expert)
10 s viewing of 2D image (for reference), following by 10 s viewing of stereoscopic image (for evaluation)
On analysing the correlation between the results of the subjective evaluation tests relating to visual comfort and the statistical quantities (mean, range, minimum, maximum, variance) of the parallax distributions, it can be seen that the parallax range exhibits a strong correlation with a correlation coefficient of −0.86 (99% confidence).
Figure 12 shows the relationship between the parallax distributions and visual comfort of the images used in the evaluation. The vertical axis shows the amount of parallax in pixel units with a value of zero corresponding to the position of the screen, and the horizontal axis shows the visual comfort derived by using the method of successive categories to combine the psychometric values. The plotted points represent the mean values of the parallax distributions in the images, and the vertical bars represent the range of the parallax distributions. From Fig. 12, we can see that in each image evaluated as being visually comfortable, the range of the parallax distribution is approximately 60 pixels or less. This translates to a value of 0.3 diopters. Images are evaluated as being comfortable when the parallax distributions are in the range from approximately 30 pixels in front of the screen to approximately 65 pixels behind it.
Next, we investigated the relationship between the average values of the parallax distributions and the visual comfort of the images. With seven of the images, we performed visual comfort evaluation tests in which the average value of the parallax distribution was shifted to different positions. These tests were performed with 20 test subjects. The other test conditions were the same as in Table 7. Figure 13 shows the experimental results. In this Figure, the points plotted with outlined symbols represent data obtained without horizontal shifting. As Fig. 13 shows, as the average value of the parallax distribution became closer to the screen position, the images were evaluated as being more visually comfortable.
Parallax distribution vs. visual comfort
Average values of the parallax distributions vs. visual comfort
When scenes are limited to small values of parallax there might be a reduction in the positive effects of the stereoscopic images, such as the sense of presence. In the tests reported in § 3, the images were evaluated in terms of their sense of presence as well as their visual comfort. Specifically, the stereoscopic images were presented with 2D images as a standard reference, and their sense of presence was evaluated on a seven-grade scale.
In the analysis of the test results, we found no significant correlation between the sense of presence scores and the statistical quantities of the parallax distributions. We extracted the images for which the stereoscopic image was evaluated as more visually comfortable than the 2D image (35 images in total), and as a result of analysing these images, we showed that there is a strong correlation between the range of the parallax distribution and the sense of presence (correlation coefficient 0.65).
On the other hand, we observed no factorial effect of the average value of the parallax distribution on the sense of presence evaluation scores.