An implementation for the "Can Human Sex Be Learned Using Only 2D Keypoint Estimations?" paper (arXiv).
Use pip3 install requirements.txt
or docker build -t sex-recognition .
.
To run the training, first download (PETA and/or 3DPeople) and prepare the datasets:
python3 src/prepare_datasets.py --name peta --dataset peta
Then you can run the training:
python3 main.py --name peta --train_datasets peta --test_dataset peta --arch fcn
To evaluate the model, run the experiments multiple times (as input data is small and the architecture is simple, it should take only few minutes per experiment):
./eval peta peta
To get the boxplots and the correlations from the paper, use the scripts from report/
directory:
python3 correlation.py
python3 report.py peta
You can also combine multiple training datasets, for example:
python3 main.py --name peta --train_datasets 3dpeople,peta --test_dataset peta --arch fcn
See more data preparation and training options by:
python3 src/prepare_datasets.py -h
python3 main.py -h
MIT