R³CNN is a general CNN architecture to integrate the relative ranking of faces in terms of aesthetics to improve the performance of facial beauty prediction.

• Caffe (compiled with pycaffe)
• python
• numpy
• matplotlib
• skimage

• Build Caffe

  make all -j16
  make test
  make pycaffe
  

• Add the python directory into the environment variables

  • Open bash file: sudo gedit ~/.bashrc
  • Add the following setence into the file: export PYTHONPATH=brl/caffe/python:$PYTHONPATH
  • Update the environment variable: source ~/.bashrc

• Dataset download:

  Our method is trained and verified on SCUT-FBP5500 benchmark dataset. The facial images should be put under examples/data/faces, where train and test set have been already provided.

• Image pairs generation:

  cd examples/data/
  python create_pair.py
  

• Mean file computation:

  sh mean.sh
  

• First stage: conventional training for ResNeXt-based regression model, using pretrained model on ImageNet (download link: https://pan.baidu.com/s/12AtCeQYuYDZtUd9jZPIo1w password: enfc):

  cd examples/first_stage
  sh train.sh
  

• Second stage:

  • Rename the caffemodel obtained in the first stage as the format of 'R2Net_hinge_iter_0.caffemodel' (download link: https://pan.baidu.com/s/1Dx3H108gCvJ71fcVg3BzjQ password: p3jk) ;
  • Put 'R2Net_hinge_iter_0.caffemodel' under examples/hinge_loss/snapshot/1;
  • Use hinge loss to train R³CNN:

    cd examples/hinge_loss
    sh train.sh
    

  • If using LSEP loss to train R³CNN, you can run examples/lsep_loss/train.sh；
  • If using other backbone networks (i.e., AlexNet and ResNet-18) to train R³CNN, you can run the codes in ./examples/other_networks;

• The trained ResNeXt-based R³CNN caffemodel is available at https://pan.baidu.com/s/1YVwKrBZS4kpNWHTRs-9qTA password: xcx7 (1.6GB)
• Run the testing file:

  cd examples/test_inference
  python test_forward.py 
  

Please cite our paper:

@article{lin2019regression,
  title={Regression guided by relative ranking using convolutional neural network (R3CNN) for facial beauty prediction},
  author={Lin, Luojun and Liang, Lingyu and Jin, Lianwen},
  journal={IEEE Transactions on Affective Computing},
  year={2019},
  publisher={IEEE}
}

For any questions, please feel free to contact Dr. Lin (linluojun2009@126.com) or Prof. Jin (eelwjin@scut.edu.cn).

This code is free to the academic community for research purpose only. For commercial purpose usage, please contact Prof. Lianwen Jin (eelwjin@scut.edu.cn).