1 1/ 13 Klepnutím lze upravit styl předlohy Klepnutím lze upravit styl předlohy Soft biometric traits in de identification process Hair Jiri Prinosil...
Klepnutím lze upravit styl předlohy Klepnutím lze upravit styl předlohy nadpisů. nadpisů. www.splab.cz
Soft biometric traits in de‐identification process ‐ Hair Jiri Prinosil Jiri Mekyska Zdenek Smekal
Brno University of Technology
2 / 13
SPLAB, VUT
Hair and hair color Klepnutím lze upravit styl předlohy nadpisů.
• Supplementary trait for identification • Hair color can be used for dividing people into sub‐groups – Assigning hair color index to a person in a video sequence – These indices can be used as filtering clue when searching in databases
Brno University of Technology
3 / 13
SPLAB, VUT
Hair segmentation method Klepnutím lze upravit styl předlohy nadpisů. • State of the art methods – static images only – simple background – high processing time
Hair segmentation method Klepnutím lze upravit styl předlohy nadpisů. • Face detection stage – position of face detector is obtained – positions of eyes and nose are obtained – Viola‐Jones detector – face tracker based on motion estimation using rectangular grid
Brno University of Technology
5 / 13
SPLAB, VUT
Hair segmentation method Klepnutím lze upravit styl předlohy nadpisů. • Background subtraction – silhouette obtained
• Head segmentation – head is obtained by using face position and silhouette
Brno University of Technology
6 / 13
SPLAB, VUT
Hair segmentation method Klepnutím lze upravit styl předlohy nadpisů.
• Skin segmentation
– the hair area is given as the difference between head and skin segment – flood‐filling approach is used – flooding starts from different seed points – positions of seed‐points are set according to positions of eyes and nose
Brno University of Technology
7 / 13
SPLAB, VUT
Hair segmentation method Klepnutím lze upravit styl předlohy nadpisů. • Skin segmentation – to suppress propagating of flood to hair area, only a compact part around a particular seed‐point is taken from the flood area – finally, a skin area is determined as a convex hull of union of selected parts around different seed points
• Hair segmentation – hair is given a the difference between hair and skin
Brno University of Technology
SPLAB, VUT
8 / 13
Hair segmentation method Klepnutím lze upravit styl předlohy nadpisů. • Results – Tested on 28 video‐sequences of avg. length of 4 seconds – Compared with ground‐truth masks Measure Precision Recall
Average
Median
0.84% 0.49%
0.92% 0.48%
– Processing times per frame Phase Face detection Background subtraction Head and skin segmentation
Average time [ms] 121 55 38
Brno University of Technology
SPLAB, VUT
9 / 13
Color estimation Klepnutím lze upravit styl předlohy nadpisů. • Experimental setup – Tested on 28 people in available video‐sequences – A color of each subject was subjectively classified into one of five categories (black, brown, blond, red, gray/white) as a reference
• Color estimation model – HSV space is equally spaced to 10 ranges for every dimension – 1000 HSV sub‐spaces are obtained – Assumption, that colors in each sub‐space are similar – Mean value of each sub‐space is subjectively classified Brno University of Technology
10 / 13
SPLAB, VUT
Color estimation Klepnutím lze upravit styl předlohy nadpisů. • Color estimation – Color of each hair pixel is evaluated using the described model – Histogram of colors in one frame is calculated – Histograms of multiple frames are summed and the biggest bin determines the hair color of a person in video‐ sequence
Brno University of Technology
SPLAB, VUT
11 / 13
Color estimation Klepnutím lze upravit styl předlohy nadpisů. • Results – Tested on 28 video‐sequences Man. extracted hair masks
Aut. extracted hair masks
Estimation
Blk.
Brw.
Bld.
Red
Wht.
Perc.
Blk.
13
1
0
0
0
92.9%
Brw.
2
5
0
0
1
62.5%
Bld.
0
0
2
0
1
66.6%
Red
0
0
0
1
0
100%
Wht.
0
0
0
0
2
100%
Subj. evaluation
Subj. evaluation
Estimation Blk.
Brw.
Bld.
Red
Wht.
Perc.
Blk.
13
1
0
0
0
92.9%
Brw.
1
6
0
0
1
75.0%
Bld.
0
0
2
0
1
66%
Red
0
0
0
1
0
100%
Wht.
0
0
0
0
2
100%
– Results are different only in one case Brno University of Technology
12 / 13
SPLAB, VUT
Experiment Klepnutím lze upravit styl předlohy nadpisů.
Brno University of Technology
13 / 13
SPLAB, VUT
Experiment Klepnutím lze upravit styl předlohy nadpisů.
Brno University of Technology
14 / 13
SPLAB, VUT
Experiment Klepnutím lze upravit styl předlohy nadpisů.
Brno University of Technology
15 / 13
SPLAB, VUT
Experiment Klepnutím lze upravit styl předlohy nadpisů.
Brno University of Technology
16 / 13
SPLAB, VUT
Klepnutím lze upravit styl předlohy Klepnutím lze upravit styl předlohy nadpisů. nadpisů. www.splab.cz
