Graphics Reference
In-Depth Information
An alternate approach to resource-constrained stream mining involves cons-
tructing topologies of classifiers based on hierarchical semantic concepts, and allow-
ing individual classifiers in the topology to operate at different performance levels
given the resources allocated to them. The performance level is determined by a clas-
sifier operating point that corresponds to the selected trade-off between probability
of detection
p
D
and probability of false alarm
p
F
. Here, the probability of detection
is defined as
p
D
=
p
tn
, where
p
tp
and
p
tn
denote, respectively, the probability
of a true positive, and the probability of a true negative.
This approach is illustrated in Fig.
12.2
, where the curve on the right side shows a
profile of the classifier accuracy in terms of the
detection error trade-off
(
DET
)—i.e.,
the trade-off of
p
D
versus
p
F
. Examples of operating points include decision thresh-
olds for likelihood ratio tests or SVM normalized scores. Hence, instead of deciding
on what fraction of the data to process, as in load-shedding approaches, such an
approach determines
how
the available data should be processed given the underly-
ing resource allocation. A solution based on this approach for configuring filtering
applications that employ binary classifier chains has been proposed [
14
,
16
-
18
].
Nevertheless, general binary tree topologies go significantly beyond linearly
cascaded classifiers by providing greater flexibility in data processing, while also
posing different challenges in terms of resource-constrained configuration. Specif-
ically, while excess load can be easily handled within the optimization framework
for a binary classifier chain, using a single operating point for each classifier in a
tree generates two output streams with a total sum output rate that is fixed. Hence,
it may not be possible to simultaneously meet tight processing resource constraints
for downstream classifiers along both output edges when using only one operating
point.
p
tp
+
12.4 Dynamic, Data-Driven ASM Systems
Building on the conceptual framework of dynamically reconfigurable topologies
of classifiers introduced in Sects.
12.1
and
12.3
, an important direction for further
work on stream mining for computer vision systems is in the rigorous integration of
Dynamic Data Driven Applications Systems (DDDAS) into all aspects of processes
for design and implementation. A significant class of future challenges for embedded
computer vision therefore involves what may be referred to as
DDDAS-enabled ASM
systems
.
12.4.1 DDDAS-Enabled ASM Systems
DDDAS is a paradigm that rigorously integrates application system modeling,
instrumentation, and dynamic, feedback-driven adaptation of model and instrumen-
tation parameters based on measured data characteristics [
13
]. DDDAS methods are
