Home

Awesome

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

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 <i>caffe</i> 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 <i>caffe_ftr.py</i>.

Structure

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) .
Datasetsizeec_mc_yec_y
Training set144x1444848
Testing set128x1284840

Training

  1. The model is trained by open source deep learning framework <i>caffe</i>.
  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.

Model100% - EERTPR@FAR=1%TPR@FAR=0.1%TPR@FAR=0Rank-1DIR@FAR=1%
A97.77%94.80%84.37%43.17%84.79%63.09%
B98.13%96.73%87.13%64.33%89.21%69.46%
C98.80%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.