Information Technology Reference
In-Depth Information
Haar
LBP
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0
50
100
150
200
0
50
100
150
200
LRD
LRP
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0
50
100
150
200
0
50
100
150
200
Fig. 2. The ROC curves on MIT+CMU dataset without suppression (full line) and
with 12 suppression classifiers (dashed line). Target miss rate
α
of the suppression
classifiers is 5 %.
PAL
MS
1
1
0.95
0.95
0.9
0.9
0.85
0.85
0.8
0.8
0.75
Haar
LBP
LRD
LRP
Haar
LBP
LRD
LRP
0.75
0.7
0.65
0.7
2
4
6
8
10
12
14
16
18
2
4
6
8
10
12
14
16
18
Fig. 3. Reduction of detection time (y-axis) when suppressing single positions in dif-
ferent horizontal distance from the classified position (x-axis). Target error of the sup-
pression classifiers is 5 %.
were used - one with less constrained poses and lower quality images and one
with easier poses and good quality images. For face detection, the harder dataset
was standard MIT+CMU frontal face detection set (CMU) and the easier was
a collection of 89 images of groups of people downloaded from the Internet. The
easy set is denoted as MS and contains 1618 faces and 142M scanned positions.
The eye detection classifiers were trained on XM2VTS 1 database and tested
on BioID 2 database (104M positions, 3078 eyes) and on a easier dataset PAL 3
(111M positions, 2130 eyes) which is similar to XM2VTS. When scanning, shift
of the window was two pixels at the base detector resolution and scale factor
was 1.2. The suppression classifiers were trained on a large set of unannotated
images containing faces.
1 http://www.ee.surrey.ac.uk/CVSSP/xm2vtsdb/
2 http://www.bioid.com/downloads/facedb/index.php
3 https://pal.utdallas.edu/facedb/
Search WWH ::




Custom Search