Age estimation using transfer learning

Age estimation is hard while the face includes different races or genders.

It also make the problem harder if the data is unbalanced, i.e. the data for black peopel is rich, while is scare for white people in Afirca.

This study is focus on how to improve the age estimation accuracy acorss different races and gender.

The primary idea is to apply tranfer learning technique which can adapt the target domain with source domain.

 

We make the assumption that people's facial feature of different race or gender hold different distribution.

They differ in two ways: one is called marginal probability, another is called conditional probability.

The former one is also refered as covariant shift, could be handled by MMD and other ways.

The later one is related to the correlations of each source domain with the target domain.

 

Here, our goal is to make better age prediction. and we apply single-source domain adaptation, rather than multiple sources.

The first step is to computer alpha, which could make the mean distribution of two domain similar.

The second step is to computer beta, which is the weight indicating importance of the test sample with source domain.

2SW-MDA, two step weighting multi domain adaptation, use two weighting process with the help of svm classifier.

LWE, local weighting ...