shixiao1997/cosonet

The source code for the paper: Yirong Mao, Ruiping Wang, Shiguang Shan, Xilin Chen. COSONet: Compact Second-Order Network for Video Face Recognition. ACCV 2018

COSONet: COmpact Second-Order Network for Video Face Recognition

This paper has been accepted as ACCV 2018 Oral [paper] [supplemental material]

Network

Requirements

PyTorch 0.4.1

Python packages: pickle, matplotlib, h5py

Datasets

The training dataset WebFace and testing dataset IJB-A are released, where the faces are detected and cropped without alignment. Our trained models are also available. You can download these data from [BaiduYun].

The overlapped subjects of WebFace and IJB-A datasets are removed while training.

Test

After unzipping the data into ./data, run test_ijba.py, and you will get the result on IJB-A of our COSONet (corresponding to the ResNet_34_COSO in Table 2)

TAR = [0.6832107 0.81946075 0.85863197 0.93930644 0.9631115 ] @FAR[0.001, 0.005, 0.01, 0.05, 0.1]

Train

To train the COSONet by yourself, you can just run train_webface_resnet34_coso.py.

Contact

If you have any question, be free to contact me. My email is yirong.maoATvipl.ict.ac.cn