I was at a supermarket buying wine and I used a self-checkout, suprisingly it didn't ask an assistant to verify my age! Then I noticed it had a camera in the bezzle. So then I wondered... could we estimate the age of a person from a picture of their face and to what degree of confidence.
In this project we attempt to estimate the age of a person from a picture of their face. We do this by training on All-Ages-Faces-Dataset and regress the age with a MSE loss. This dataset has a heavy racial bias, and impacts the models performance.
This project uses conda to manage its enviroment; once conda is installed we create the enviroment and activate it,
conda env create -f enviroment.yml
conda activate age_regression. On windows; powershell needs to be initialised and the execution policy needs to be modified.
conda init powershell
Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser. This repo uses git-lfs to store the models, make sure the git-lfs files have been pulled using,
git lfs pullTo run age prediction on a directory of images run,
python evaluate_images.py --model pretrained/model_age_regression_resnext101_20.pth --images ~/code/datasets/faces/valWhen we do a joint-plot between predicted and ground-truth age we can see the model has a strong correlation between the two but there are the occasional outliers.
We can see the dataset has a large age inbalance, mostly featuring samples from people 25~35. As we're trying to maximise our performance on this dataset its has been ignored. In the future, this dataset should be resampled before training to give a uniform age distribution.
When we plot the MAE at each age group. We can see our age group with the smallest MAE is 30-35 which corresponds to our predominant age-group in the dataset. Where we have very little data we can see larger prediction errors. Its possibly interesting that the groups for ages <15 also have lower MAE, possibly because they're easier cases.
