Graphics Reference
In-Depth Information
Table 12.1
Average attempts (with the oracle bound given in parentheses) to obtain a recognition
rate of 0
.
9with2D-PCA
Method
Iteration
T
=
50
T
=
100
T
=
250
T
=
1,000
CBF
3.3 (1.7)
3.1 (1.6)
2.4 (1.5)
1.9 (1.5)
CBF no context
3.1 (1.7)
2.8 (1.6)
2.6 (1.6)
2.4 (1.6)
Q-learning
3.5 (1.7)
2.8 (1.6)
2.7 (1.5)
2.2 (1.5)
12.5.5 Numerical Results
The CBF has been evaluated by simulation. The simulation environment comprises
four mobile devices connected via an IEEE 802.11 WLAN to a cloud-computing
server. Videos of human faces are produced by random images of persons taken
from the extended Yale Face Database B (39 cropped faces of human subjects under
varying illumination) [
20
]. Each video comprises 34 images from the same person,
and is compressed to a wide range of bitrates via the H.264/AVC codec (x264 codec,
crf
). The 2D PCA algorithm [
43
] is used at the cloud side
for face recognition from each decoded video (with the required training done offline
as per the 2D PCA setup [
43
]). More than 80% of the video frames have to match to
the same person in the database to declare a given video as “recognized”. There is a
time window set for recognition, which limits the number of frames received by the
cloud under varying WLAN contention levels (delay is increased under contention
due to the backoff and retransmissions of IEEE 802.11 WLANs). Similarly, because
of randomly varying congestion in the cloud, only a limited number of the received
video frames is actually used by 2D PCA, thereby affecting the recognition rate.
Table
12.1
presents the average number of retries performed per recognition action
by the CBF method (with and without using the cloud congestion information as
context) in order to achieve a recognition rate of 90%. Results are also presented in
the following ways.
•
∈ {
4
,
14
,
24
,
34
,
44
,
51
}
An optimal solution that selects the transmission setting yielding the highest
expected recognition rate [
2
]. This solution is defined as the
oracle solution
, since
it assumes that all conditions for each case are precisely known beforehand.
•
Q-learning [
36
,
44
], as discussed in Sects.
12.5.3
and
12.5.4
.
The results indicate that after 250 recognition attempts (each attempt comprises
the retries listed), the CBF method approaches the oracle bound, and for the same
recognition rate, incurs less retries per attempt in comparison to Q-learning.
