by nanospeck on 12/7/16, 8:35 AM with 2 comments
Since I don't have any mentors on this area. I am very interested to know atleast what this specific problem area is called in Machine Learning. Are there currently any frameworks you know that can do this?
by webmaven on 12/7/16, 10:02 AM
I am not aware of any work that checks how accurate such age progression (or regression) transforms are, but there are many papers that use the realism (as evaluated by humans) of the results as a criteria.
As one example, here is a recent paper (Deep Feature Interpolation for Image Content Changes) that uses age progression as one of the evaluation tasks: https://arxiv.org/abs/1611.05507
Regardless of the model you choose, if accuracy is your goal your training data may need 2+ images of each subject at different ages, labelled as to age (and possibly year). You might not necessarily need to associate different images of the same person with each other, you just need to give your system the opportunity to learn more than just how a generic 16yo ages to a generic 37yo, etc.
As a bonus for having the images labelled by year, the same model should also be able to transform a 1972 18yo into a 2012 18yo, etc.
by billconan on 12/7/16, 4:08 PM
But I see 2 difficulties. One is collecting this data. Maybe if you have access to a passport photo database, things will be easier.
Second is how to design the cost function. direct comparing pixels doesn't seem to be the right way.
a generative model sounds better.