Information Technology Reference
In-Depth Information
Fig. 1.
Scanning an image in ordinary line-by-line fashion while using neighborhood
suppression
Table 1.
The benefit of neighborhood suppression for different features and datasets.
ROCA is the percentage difference between area under ROC without and with area
suppression. Time represents average number of features computed per position rel-
ative to the original detector without neighborhood suppression. ”single” stands for
suppressing single position. ”12” stands for suppressing 12 positions with 12 suppres-
sion classifiers. Target error of the suppression classifiers was 5 %.
Haar
LBP
LRD
LRP
dataset
value
single
12 single
12 single
12 single
12
ROCA (%) -0.02 0.07 -0.48 -3.44 -0.16 -1.08 -0.24 -2.04
Time
BioID
0.96 0.68
0.78 0.33
0.92 0.54
0.82 0.37
ROCA (%) -0.00 -0.39 -0.08 -0.21 -0.09 -0.85 -0.05 -0.44
Time
PAL
0.96 0.71
0.77 0.31
0.91 0.51
0.82 0.36
ROCA (%) -0.03 -0.36 -0.27 -1.92 -0.02 -0.49 -0.08 0.01
Time
CMU
0.93 0.62
0.74 0.31
0.93 0.62
0.87 0.47
ROCA (%) -0.04 -0.54 -0.21 -1.02 -0.02 -0.27 -0.06 -0.65
Time
MS
0.93 0.60
0.73 0.29
0.93 0.60
0.87 0.45
it is easy to fulfill as the maximum possible value of each portion of the register
can be calculated and predicted.
The suppression itself can be handled by binary mask covering positions to
be scanned. The positions marked as suppressed are then excluded from further
processing. The scanning order can remain the same as in ordinary scanning
window approach, even though it restricts the positions which can be suppressed
to those which are to the left and bottom of the currently classified position (see
Figure 1). Possibly, more e
cient scanning strategies can be developed, but such
strategies are beyond the scope of this paper.
3 Experiments and Results
We tested the neighborhood suppression approach presented in the previous text
on frontal face detection and eye detection. In both task, two separate test sets
Search WWH ::
Custom Search