Introduction
Humans sense depth by using two distinct perspective views from our left and right eyes. The visual cortex combines the pictures two eyes provide in order to perceive how far away objects are.
Based on this, the visual cortex deduces the 3D structure of the scene being observed. In computer vision, this concept is typically called stereo, where the stereo disparity is the difference in position between the two images.
In the context of visual interaction, stereo-multiplexing is the process in which two images are placed in the same place. This is a commonly used concept in 3D movie production. By having two slightly different pictures presented to viewers on a screen, a 3D depth effect can be created.
It is also used in VR technology by displaying slightly different perspectives to the left and right eye individually.
Hypothesis
"Higher stereo disparity increases viewer's ability to perceive depth differences".
Solution
An experiment will be conducted in a VR application where testers are presented with
different stimuli having different combinations of stereo disparity levels and depth differences.
Expected Outcome
The tester will be able to more accurately determine differences in depth in the
scenes with higher stereo disparity.
Methods/Process
-
Development of VR Mobile Application in Unity3D and using Google CardBoard.
-
Conducting the experiment and collecting Data.
-
Statistical Analysis.
-
Descriptive Analysis
-
Chi-Square Test
-
Development of Mobile Application
A mobile application was developed in Unity using the Google VR SDK and an experiment was conducted to test whether users have a good depth perception on two different stereo disparity levels (0.064m & 0.128m). The experiment was held on two mobile platforms of Android and iOS alone with a Google Cardboard VR headset
Experimental Setup
To simulate the stereo disparity effect, the scene in Unity reflected on the same proportion of objects in real life that 1 Unity Unit represents 1 meter in real world. The objects used in the experiment are unit cubes (1m * 1m * 1m) provided by Unity.
Through the pre-testing of the group member, the experimental scene is placed according to the image below ( figure 2 ) to have a considerable amount of 3D effects to be observed where objects are placed on the depth of 3 meters away from the camera and have an interval distance of 0.2 meters (Stereo effects are only effective in a near visual field around 3-5 meters).
The Blue line represents the facing direction of the head, and the Green & Red lines represent the look-at direction of two eyes. To have a better stereo observation, all cubes in the scene are textured so that users may notice more stereo details on the objects.
Note, in order to lower the variance of the observation, the degree of freedom is strictly limited that during the experiment, the virtual cameras are unable to rotate and move. In addition, the skybox is turned off and the background colour is set to black, so testers could focus the observation on the depth of objects in the scene.
Experiment Steps
After the application is launched, the user needs to select a combination of stereo disparity and offset in the start menu and click the start button to start the experiment. There are two stereo disparity options and 5 offset options which total 10 pairs of combinations.
01 : A scene is presented with 3 different color boxes(Red, Yellow and Green). All boxes will be first placed at a default depth of 3 meters. Then, one or zero of the boxes will be moved closer to the camera of the number of offset selected. The tester’s task is to determine which box is at a different depth than the rest (Which box is closer to viewing point) or all boxes are at the same depth. Therefore, there will be four kinds of results: RED, YELLOW, GREEN or SAME.
02 : The user keeps track of their estimated results by themselves. Then users need to click the button on Google Cardboard to get the correct answer in one of the four results (RED, YELLOW, GREEN or SAME) and compare the answer with their previous estimation (RIGHT or WRONG).
03 : User records the comparison result and clicks the button again to get the next test and repeat step 2 & 3 a total of 20 times.
A complete test consists of 20 stimuli and 10 combinations; 200 data points. Each of the group members completed a test and recorded their results which resulted in a total of 60 samples for each combination
Data Analysis
All data collected from the test is imported into a data analysis application called Jamovi to get further insight of the result. A statistical test will be conducted to verify the statistical significance of the data sets.
Results
For all 10 pairs of stereo separation and object offset combination, 40 samples are collected. In the table below, stereo separation (meter) represents the distance between two virtual cameras, object offset (meter) represents the distance in Z axis unit cube moves towards head position and 0 & 1 represents whether the estimation result is correct or incorrect.
Analysis
In the descriptive analysis, the indicated result is pretty clear that in all groups of 0.128 meters stereo separation, the observed depth correctness is higher than the corresponding groups of 0.064 meters stereo separation.
When the offset stereo separation number is the same and only takes the object offset number as the variable, we can see a constant increase of correctness when the offset number is increasing. It is understandable that the depth difference is bigger, the more obvious the user may be able to sense the depth.
When the offset number is lower, the effect is more obvious that in the offset of 0.025 the correctness increased more than 10% which is the highest in all 5 offset groups. In addition, when the offset number reaches 0.125, the result does not show a significant difference in the correctness where in both 0.064 meters and 0.128 meters groups, the correctness stays 96.7%. We believe it is due to the depth difference being too obvious that the user does not need the contribution from the stereo separation difference to sense the depth.
In order to verify the statistical significance of the data sets, our group performed three Chi-square tests on the relationship of object offset and observation correctness when the stereo separation is 0.064 meters, 0.128meters and combined.
The significance level (a < 0.01) chosen for the test is 0.01 which is a commonly used value in mathematical statistical analysis.
The null-hypothesis (H0) is:
Object offset value is independent to the estimation correctness when stereo separation is 0.064 or 0.128 or combined.
The Alternative hypothesis (H1) is: Object offset value is associated with the estimation correctness when stereo separation is 0.064 or 0.128 or combined.
The result of the three groups of Chi-square tests all showed the p value lower than 0.001 which is lower than the chosen level of significance. Meaning, we are able to reject the H0 and accept the H1 that proves the data sets are under statistical significance.
In conclusion, the result from the experiment is promising that a certain degree of higher stereo separation can increase user’s ability to sense depth in a 3D environment.
Discussion
The results show that higher stereo disparity does increase a viewer’s ability to perceive depth. However, we believe this experiment can be improved and this topic requires further research.
In the experiment, each stimuli consists only of 3 cubes; all of which are at a similar distance from the viewer and each other. In order to further investigate the hypothesis, a more advanced 3d environment could be utilized. New environments with different size objects, different distances between objects or more complex objects could improve the experiment. Therefore, the methods we used in the current experiment also need to be improved to adapt to a more complicated scene.
A potential weakness of the experiment is that it did not include a time limit to make judgements. This can be problematic because it leaves it up to the tester to try to be consistent. Additionally, the testing was conducted with different cardboard viewers and mobile devices. The mobile devices had different screen resolutions and pixel densities. These weaknesses may or may not have had an effect on the results.
Future research into the hypothesis should utilize more stereo disparity values and explore its limits. Our experiment only used two stereo disparities, 0.064 and 0.128 meters. We started with 0.064m because it is the default separation the Google VR SDK utilizes.
We believe that the hypothesis might not hold at very high disparities and depth perception will worsen if the eye cameras get to be too far apart because the images produced by the cameras will be too dissimilar.
Therefore, the relationship between stereo disparity and depth perception could first follow a positive relationship then when the separation is too big, the user can no longer merge two viewing images into one which decreases the depth sensing ability.