Information Technology Reference
In-Depth Information
5
Experimental Results
Experimental Environment
1.
CPU: Intel ® Core TM i5-650 3.2GHz processor
2.
Memory: 12GB
OS: Windows 7 TM
3.
4.
Programming Language: Matlab, C++, Python.
We use FERET database [22] as the data set for our age and gender estimation algo-
rithm. All the images in the test set are 512x768 pixels. The data set contains 2722
grey-scale pictures: 1007 female pictures and 1715 male pictures. We use 5-fold cross
validation where each fold would be trained on 2178 images and tested on 544 im-
ages.
5.1
The Effect of Shifting on Gender Estimation
Table 2. The accuracy of the gender estimation on different regions separately using different
amounts of shift in pixels
0 pixels
30 pixels
60 pixels
90 pixels 120 pixels
Left Eyebrow
78.54% 78.46% 82.93% 81.53%
74.70%
Right Eyebrow
78.20%
80.62%
82.65%
84.09%
76.53%
Left Eye
71.94% 76.90% 82.46% 87.75%
82.76%
Right Eye
72.42% 74.43% 83.55% 86.47%
85.79%
Nose
86.61%
71.91% 73.83% 79.91% 85.59%
Mouth
78.12% 81.40% 86.87% 87.53%
84.87%
Face
86.60% 87.29% 88.42% 80.91%
66.73%
Voting
89.12%
88.95%
90.91%
91.05%
85.58%
We expected the best results would be in the region of 10 to 20 pixels shift, this way
all the patches would overlap with the original patch (Fig. 5). However, as we keep
increasing the shifting amount, the results keep getting better and would only drop off
when the patches would start shifting to the outside of the face regions. The cause of
this is that all areas of the face carry information about gender and age. Moreover,
when there is less overlapping between regions, we can extract more information that
can lead to better estimations.
Search WWH ::




Custom Search