We conduct the experiments on MORPH
Album 2 database, which is the largest aging database we know. For equal
comparison and reproducible research, we supply the detailed evaluation
protocols and 21 facial landmarks for all face images.
MORPH Album 2 [1] contains about 55,000 face images of more than 13,000 subjects. The capture time spans from 2003 to 2007. Age ranges from 16 to 77 years. Although it is a good and large database, the distributions of gender and ethnicity are uneven. The Male-Female ratio is about 5.5 : 1 and the White-Black ratio is about 4 : 1. Except for White and Black, the proportion of other ethnicity is very low (4%).
To
use the database effectively, we follow the previous way [2] to pre-processing
the database and split it into three non-overlapped subsets S1, S2 and S3 randomly.
First, all images in MORPH are processed by a face detector. Because MORPH
contains some non-face images (e.g., tattoo), they are removed from the
database after this step. The number of face images in the processed database
is 55244. Then the facial landmarks of face images are localized by ASM and
local aligned patches are cropped based on the landmarks described in Section
3.1. We construct the S1, S2 and S3 subsets by two rules: 1) Making Male-Female
ratio to 3; 2) Making White-Black ratio to 1.
The information of the subsets
is shown in Table 1. In all experiments, the training and testing are repeated
in two times: 1) training on S1, testing on S2+S3 and 2) training on S2, testing on S1 + S3. The performance of the two experiments and their
average are reported. For age and gender estimation, all images in Table 1 are
used. For ethnicity classification, the images in “Other” column are neglected.
Table 1. The information of
the pre-processed MORPH Album 2 and S1, S2, S3 subsets.
|
Black |
White |
Other |
||||
Male |
S1: 4012 |
S2: 4012 |
S3: 28835 |
S1: 4012 |
S2: 4012 |
S3: 0 |
S3: 1845 |
Female |
S1: 1305 |
S2: 1305 |
S3: 3166 |
S1: 1305 |
S2: 1305 |
S3: 0 |
S3: 130 |
The file list of
Exp1: S1, S1 Test (S2 + S3) and
Exp2: S2, S2 Test (S1 + S2) can be
downloaded here.
The format of the list file is:
[filename] [age] [ethnicity] [gender],
where
[age] is between 16 and 77,
[ethnicity] is 0 for Black, 1 for
White, -1 for others
[gender] is 0 for Male, 1 for Female
Some sample lines are shown as follows,
Album2/009055_0M54.jpg 54 1 0
Album2/019066_08M70. jpg 70 0 0
Album2/262418_01F20. jpg 20 0 1
Album2/261887_00M26. jpg 26 -1 0
Album2/268066_00F36. jpg 36 -1 1
…
Note (2015-8-28): After the publication of this work, we found that some face images in S2 and S2_Test are overlapped, therefore we re-generate a test set "S2_Test_Fixed" to fix this problem. If you want compare to our paper, you should use S2_Test. If you want evaluate the performance more objective, you should use S2_Test_Fixed. We feel very sorry for our mistake. Many thank Alberto Escalante (alberto.escalante@ini.rub.de) for pointing out this problem.
In our paper, we use 21 facial landmarks to crop aligned patches from face images. The positions of the landmarks are shown in Fig. 1.
Figure 1. The 21 facial landmarks used in the experiments.
The facial landmarks can be downloaded here. The landmarks are detected by a multi-view ASM automatically. Maybe there are some errors in the coordinates, but we just leave them there to test the robustness of our convolutional network.
The
landmarks for each face image are save to an independent text file, e.g., the
landmarks of “009055_0M54.jpg”
is in “009055_0M54.jpg.shape21”. The format of “shape21” file is:
x1, y1
x2, y2
x3, y3
…
x21, y21
The article based on this evaluation protocols has been published in ACCV 2014. Please cite the following paper in any published work if you use the protocols and data.
[1] Rawls, A., Ricanek, K.: Morph: Development and
optimization of a longitudinal age progression database. In: COST 2101/2102
Conference. (2009) 17-24.
[2] Guo, G., Mu, G.: Human age estimation: What is
the influence across race and gender? In: Computer Vision and Pattern
Recognition Workshops (CVPRW), 2010 IEEE Computer Society Conference on. (2010)
71-78.