In this task the goal is to estimate the age of the persons depicted in the individual images.
We propose to use the following evaluation metrics:
- Mean Absolute Error (MAE)
- Cumulative Score (CS)
- MAEs per decade (MAE/D)
- Average over MAEs per year (AMAE/y)
The label/fold files are structured such that there is one line per image. Each line starts with the filename followed by the fold ID and the age in years. Each of the values is separated by a tab. An example line looks like the following:
224896_00M25.JPG 0 25
where 224896_00M25.JPG is the filename of the image, 0 is the fold ID, and 25 shows that the age of the person depicted on the image is 25 years.
We propose to use the MORPH-II database for the controlled labratory condition also for age estimation.
Due to the size of the MORPH-II database the 5-fold cross-validation evaluation scheme shall be used for the evaluation of the approaches for the controlled condition. To prevent algorithms from learning the identity of the persons in the training set rather than the age it has to be made sure that all images of individual subjects are only in one fold at a time. Additionally, the folds are selected in such a way that the distribution of age, gender and ethnicity in the folds is similar to the distribution in the whole database. The file list for these folds can be found in the Downloads section below.
The Face and Gesture Recognition Research Network (FG-NET) aging database is proposed to be used for the uncontrolled real-life condition.
For the evaluation of the approaches for the uncontrolled condition the leave-one-person-out (LOPO) evaluation scheme shall be used, since this is a common evaluation scheme for this database, due to the small size of the FG-NET Aging database. For LOPO, all samples of a single person are used for testing and the remaining samples for training. This is done for all subjects so that each person is once used for testing, resulting in 82 folds. This evaluation scheme makes sure that images of a person are not in the testing and training set at the same time, so that the classifier cannot learn some "intra personal" relations. The file list for these folds can be found in the Downloads section below.
- Complete Evaluation Guidelines
- Folds for controlled condition of age estimation (old MORPH distribution)
- Folds for controlled condition of age estimation (new MORPH distribution; thanks to Nesli Erdogmus)
- Folds for controlled condition of age estimation on balanced dataset
- Folds for uncontrolled condition of age estimation
- Folds for uncontrolled condition of age estimation on balanced dataset
- Matthias Steiner, Karlsruhe Institute of Technology, Germany
- Tobias Gehrig, Karlsruhe Institute of Technology, Germany
|FG-NET Aging Database
K. Ricanek Jr. and T. Tesafaye
|MORPH: A Longitudinal Image Database of Normal Adult Age-Progression
In IEEE 7th International Conference on Automatic Face and Gesture Recognition (FGR’06), Southampton, UK, Apr. 2006, pp. 341–345.
Age Group Estimation
In this task the goal is to estimate the age group (or category) of the persons depicted in the individual images.
As dataset the Images of Groups Dataset is proposed. The dataset consists of Flick face images labeled with 7 age categories.
Some results on this data were published in C. Shan. "Learning Local Features for Age Estimation on Real-life Faces", ACM International Workshop on Multimodal Pervasive Video Analysis, Florence, Italy, 2010.
- Caifeng Shan, Philips Research, Netherlands
A. Gallagher and T. Chen
|Understanding images of groups of people
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 256-263, 2009.
|Learning Local Features for Age Estimation on Real-life Faces
ACM International Workshop on Multimodal Pervasive Video Analysis, Florence, Italy, 2010.