IEEE Consumer Electronics Magazine - May 2018 - 29

Scene Corner Location in y Direction

Scene Corner Location in x Direction
0
200
400
600
800

Poorer identification is also the
case with dark-colored food, items
with a rough surface, foods of
similar color, and food that is
partially occluded.

100
200
300
400
500
600
700
800

Scene Corner Location in y Direction

FIGURE 6. The jittering variation with the ORB descriptor. The
coordinates of all four scene corners for each frame are shown.
The blue points represent the location of the scene corners
acquired using a simple background (we used pure white), while
the red points represent the location of the scene corners
acquired using a real scene with a complicated background. The
points in the rectangle represent valid scene corners, because the
Google Glass resolution is 640 × 360 pixels.

Scene Corner Location in x Direction
0
100 200 300 400 500 600 700

100
200
300
400
500

and thus more identifiable keypoints, in a real scene with a
complicated background than there are in a scene with a simple background. Even with a pure white background, the
average jittering level of the proposed system can still be
limited to approximately 14.20 pixels (<4% of the screen
height or <2.3% of the width). Under the same conditions,
ORB could also achieve a 0.12-pixel jittering level in the
case of a complex background, which is better than LDB.
However, with a pure white background, the average jittering level increased to 23.63 pixels, and more outliers existed, as shown in Figure 6.
For measuring relative distances, we moved the device
back and forth slightly, while rotating it to artificially create
a somewhat noisy scenario. In Figure 8, the results of the
two implementations, LDB and ORB, are displayed and
compared. During the process of moving and rotating the
Google Glass, the change in distance between subsequent
frames was relatively stable. Notably, although jitter was
always present to some extent, there were no jumps of more
than 20% of the original value, even though the device was
rotated over 180°. Note that the large jitter at the beginning
was caused by human disturbance. In this experiment, the
performance of LDB and ORB was comparable. The result
also revealed that the proposed system remained stable while
tracking multiple objects.

TRADEOFF BETWEEN JITTERING AND RIS

600

FIGURE 7. The jittering variation with the LDB descriptor. All of
the settings are the same as in Figure 5.

Our goal was to achieve a high level of performance with
both AR and RIS. However, this goal must be tempered by
the fact that high performance with one element often implies
reduced performance with another. As shown in Figure 9, the

Distance (Pixels Unit)

350
ORB
LDB

300
250
200
150
100

100

150

200

250

300

350
400
450
Frame Number

500

550

600

650

700

750

800

FIGURE 8. The relative distance of two food items between subsequent frames.

MAy 2018

IEEE Consumer Electronics Magazine

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May 2018