Hyper Face implementation which predicts face/non-face, landmarks, pose and gender simultaneously.

This is NOT official implementation.

This software is released under the MIT License, see LICENSE.txt.

• Chainer implementation
• Image viewer on web browsers

• Python 2.7
• Chainer 1.14.0
• OpenCV 2.4.9
• Flask 0.11.1
• Flask_SocketIO 2.4
• Dlib 19.1.0

• Python 3.5
• Chainer 1.14.0
• OpenCV 3.1.0
• Flask 0.10.1
• Flask_SocketIO 2.2
• Dlib 19.1.0

Important variables are configured by config.json.

Set gpu positive number to use GPU, port numbers of web servers and so on.

Download AFLW Dataset and AlexNet Caffe Model, expand them and set aflw_sqlite_path, aflw_imgdir_path, and alexnet_caffemodel_path in config.json

Pre-training with RCNN_Face model.

python ./scripts/train.py --pretrain

Open http://localhost:8888/, http://localhost:8889/ and http://localhost:8890/ with your web browser to see loss graphs, network weights and predictions. Port numbers are configured by config.json.

python ./scripts/train.py --pretrainedmodel result_pretrain/model_epoch_40

Use arbitrary epoch number instead of 40.

To skip training, please use trained model from here (or here (Do not expand as zip)).

python ./scripts/use_on_test.py --model model_epoch_190

Open http://localhost:8891/ to see predictions.

Set your image file with --img argument. The dependence are less than other tests and demos.

python ./scripts/use_on_file.py --model model_epoch_190 --img sample_images/lena_face.png

Input images are contained in sample_images directory.

Open http://localhost:8891/ to see demos.

python ./scripts/demo_on_test.py --model model_epoch_190

Demo using AFLW test images

python ./scripts/demo_live.py --model model_epoch_190

• Tune training parameters.
• Fix pose drawing.
• Implement post processes.
• Tune post processes parameters.