A Light CNN for Deep Face Representation with Noisy Labels

Citation

If you use our models, please cite the following paper:

@article{wulight,
  title={A Light CNN for Deep Face Representation with Noisy Labels},
  author={Wu, Xiang and He, Ran and Sun, Zhenan and Tan, Tieniu}
  journal={arXiv preprint arXiv:1511.02683},
  year={2015}
}
@article{wu2015lightened,
  title={A Lightened CNN for Deep Face Representation},
  author={Wu, Xiang and He, Ran and Sun, Zhenan},
  journal={arXiv preprint arXiv:1511.02683},
  year={2015}
}
@article{wu2015learning,
  title={Learning Robust Deep Face Representation},
  author={Wu, Xiang},
  journal={arXiv preprint arXiv:1507.04844},
  year={2015}
}

Updates

• Dec 16, 2016
  • The MS-Celeb-1M clean list is uploaded: Baidu Yun, Google Drive.
• Nov 08, 2016
  • The prototxt and model C based on caffe-rc3 is updated. The accuracy on LFW achieves 98.80% and the TPR@FAR=0 obtains 94.97%.
  • The performance of set 1 on MegaFace achieves 65.532% for rank-1 accuracy and 75.854% for TPR@FAR=10^-6.
• Nov 26, 2015
  • The prototxt and model B is updated and the accuracy on LFW achieves 98.13% for a single net without training on LFW.
• Aug 13, 2015
  • Evaluation of LFW for identification protocols is published.
• Jun 11, 2015
  • The prototxt and model A is released. The accuracy on LFW achieves 97.77%.

Overview

The Deep Face Representation Experiment is based on Convolution Neural Network to learn a robust feature for face verification task. The popular deep learning framework caffe is used for training on face datasets such as CASIA-WebFace, VGG-Face and MS-Celeb-1M. And the feature extraction is realized by python code caffe_ftr.py.

Structure

• Code
  • data pre-processing and evaluation code
• Model
  • caffemodel.
    • The model A and B is trained on CASIA-WebFace by caffe-rc.
    • The model C is trained on MS-Celeb-1M by caffe-rc3.
• Proto
  • Lightened CNN implementations by caffe
• Results
  • LFW features

Description

Data Pre-processing

1. Download face dataset such as CASIA-WebFace, VGG-Face and MS-Celeb-1M.
2. All face images are converted to gray-scale images and normalized to 144x144 according to landmarks.
3. According to the 5 facial points, we not only rotate two eye points horizontally but also set the distance between the midpoint of eyes and the midpoint of mouth(ec_mc_y), and the y axis of midpoint of eyes(ec_y) .

Dataset	size	ec_mc_y	ec_y
Training set	144x144	48	48
Testing set	128x128	48	40

Training

1. The model is trained by open source deep learning framework caffe.
2. The network configuration is showed in "proto" file and the trained model is showed in "model" file.

Evaluation

1. The model is evaluated on LFW which is a popular data set for face verification task.
2. The extracted features and lfw testing pairs are located in "results" file.
3. To evaluate the model, the matlab code or other ROC evaluation code can be used.
4. The model is also evaluated on MegaFace. The dataset and evaluation code can be downloaded from http://megaface.cs.washington.edu/

Results

The single convolution net testing is evaluated on unsupervised setting only computing cosine similarity for lfw pairs.

Model	100% - EER	TPR@FAR=1%	TPR@FAR=0.1%	TPR@FAR=0	Rank-1	DIR@FAR=1%
A	97.77%	94.80%	84.37%	43.17%	84.79%	63.09%
B	98.13%	96.73%	87.13%	64.33%	89.21%	69.46%
C	98.80%	98.60%	96.77%	94.97%	93.80%	84.40%

The details are published as a technical report on arXiv.

The released models are only allowed for non-commercial use.