SSPP-DAN-TensorFlow

Tensorflow implementation of SSPP-DAN: Deep Domain Adaptation Network for Face Recognition with Single Sample Per Person

Prerequisites

• Python 3.5.2
• Tensorflow 1.12.0-gpu-py3
• OpenCV 4.1.1
• SciPy

We recommend the following instuctions.

• Pull docker image (docker pull tensorflow/tensorflow:1.12.0-gpu-py3)
• In the docker container apt-get update pip install scikit-image apt-get install -y libsm6 libxext6 libxrender-dev pip install opencv-python

Usage

First, download the dataset or the pickle files that we already generated. After all pickle files are download, move them into the 'SSPP-DAN/data/eklfh_pkl' folder.

Directory Tree

|-- DAN.py
|-- README.md
|-- data
|   |-- EK-LFH
|   |-- SCface
|   |-- __init__.py
|   |-- data_manager.py
|   |-- eklfh_pkl
|   |   |-- eklfh_s1_tgt_test.pkl
|   |   |-- eklfh_s1_tgt_train.pkl
|   |   |-- eklfh_s2_tgt_test.pkl
|   |   |-- eklfh_s2_tgt_train.pkl
|   |   |-- eklfh_src_train_ori.pkl
|   |   |-- eklfh_src_train_ori_3D.pkl
|   |   |-- eklfh_src_train_ori_3D_semi.pkl
|   |   |-- eklfh_src_train_ori_semi.pkl
|   |-- pkl_generate_eklfh.py
|   |-- pkl_generate_scface.py
|-- pretrained
|   |-- VGG_Face.py
|   |-- __init__.py
|   |-- get_vggface.sh
|-- test_model.py
|-- train_model.py
|-- util
    |-- Logger.py
    |-- OPTS.py
    |-- PyMatData.py
    |-- __init__.py
    |-- flip_gradient.py
    |-- img_proc.py

Then run get_vggface.sh in the SSPP-DAN/pretrained folder to use the pre-trained VGG-Face model.

To train a model with downloaded dataset:

$ python train_model.py --dataset='eklfh_s1' --exp_mode='dom_3D' 

To test with an existing model:

$ python test_model.py --dataset='eklfh_s1' --exp_mode='dom_3D'  --summaries_dir 'exp_eklfh_s1/tuning/exp_2_dom__batch_64__steps_10000__lr_2e-05__embfc7__dr_0.3__ft_fc7' 

Results

Facial feature space (left) and its embedding space after applying DA (right). The subscript “s” and “t” in the legend refer to the source and target domains, respectively.

Author

Sungeun Hong e: [email protected] w: www.csehong.com