ten face doc

DocFace: Matching ID Document Photos lớn Selfies

By Yichun Shi and Anil K. Jain

Bạn đang xem: ten face doc

Update Notes

  • 09/28/2018: Added a phác thảo code for face alignment.
  • 09/20/2018: The nội dung of DocFace+ is added, including DIAM-Softmax, partially shared sibling networks and tư vấn for multiple ID/selfie per class.


  1. Introduction
  2. Citation
  3. Requirements
  4. Usage
  5. Models
  6. Results


This repository includes the TensorFlow implementation of DocFace and DocFace+, which is a system proposed for matching ID photos and live face photos. DocFace is shown lớn significantly outperform general face matchers on the ID-Selfie matching problem. We here give the example training code and pre-trained models in the paper. For the preprocessing part, we follow the repository of SphereFace lớn align the face images using MTCNN. The user can also use other methods for face alignment. Because the dataset used in the paper is private, we cannot publish it here. One can test the system on their own dataset.


@article{shi2018docface+, title = {DocFace+: ID Document lớn Selfie Matching}, author = {Shi, Yichun and Jain, Anil K.}, booktitle = {arXiv:1809.05620}, year = {2018} } @article{shi2018docface, title = {DocFace: Matching ID Document Photos lớn Selfies}, author = {Shi, Yichun and Jain, Anil K.}, booktitle = {arXiv:1805.02283}, year = {2018} }


  1. Requirements for Python3
  2. Requirements for Tensorflow r1.2 or newer versions.
  3. Run pip install -r requirements.txt for other dependencies.
  4. Requirements for Matlab 2014b and Caffe for MTCNN face alignment.


Part 1: Preprocessing

1.1 Dataset Structure

Download the Ms-Celeb-1M and LFW dataset for training and testing the base model. Other dataset such as CASIA-Webface can also be used for training. Because Ms-Celeb-1M is known lớn be a very noisy dataset, we use the clean list provided by Wu et al. Arrange Ms-Celeb-1M dataset and LFW dataset as the following structure, where each subfolder represents a subject:

Aaron_Eckhart Aaron_Eckhart_0001.jpg Aaron_Guiel Aaron_Guiel_0001.jpg Aaron_Patterson Aaron_Patterson_0001.jpg Aaron_Peirsol Aaron_Peirsol_0001.jpg Aaron_Peirsol_0002.jpg Aaron_Peirsol_0003.jpg Aaron_Peirsol_0004.jpg ...

For the ID-Selfie dataset, make sure all the foldesr in such a structure, where ID images and selfies start with "A" and "B", respectively :

Subject1 A001.jpg B001.jpg B002.jpg Subject2 A001.jpg B001.jpg ...

1.2 Face Alignment

To ensure the peroformance, we align all the face images using the original MATLAB version of MTCNN following SphereFace. Here, we provide a simpler code lớn align a given dataset thư mục. To use the code, you need lớn install Caffe for Matlab, and clone the repo of MTCNN and Pdolloar. Then fill in their paths in the following lines in align/face_detect_align.m:

imglist = importdata('/path/to/input/imagelist.txt'); output_dir = '/path/to/output/dataset'; ... matCaffe = '/path/to/caffe/matlab/'; pdollarToolbox = '/path/to/toolbox'; MTCNN = '/path/to/mtcnn/code/codes/MTCNNv1'; ... modelPath = '/path/to/mtcnn/code/codes/MTCNNv1/model';

Run the following command in Matlab for face alignement:

run align/face_detect_align.m

Part 2: Training

Note: In this part, we assume you are in the directory $DOCFACE_ROOT/

2.1 Training the base model

  1. Set up the dataset paths in config/basemodel.py:

    # Training dataset path train_dataset_path = '/path/to/msceleb1m/dataset/folder' # Testing dataset path test_dataset_path = '/path/to/lfw/dataset/folder'

  2. Due lớn the memory cost, the user may need more than thở one GPUs lớn use a batch size of 256 on Ms-Celeb-1M. In particular, we used four GTX 1080 Ti GPUs. In such cases, change the following entry in config/basemodel.py:

    # Number of GPUs num_gpus = 1

  3. Run the following command in the terminal:

    python src/train_base.py config/basemodel.py

    After training, a model thư mục will appear underlog/faceres_ms/. We will use it for fine-tuning. If the training code is lập cập more than thở once, multiple folders will appear with time stamps as their names. The user can also skip this part and use the pre-trained base model we provide.

2.2 Fine-tuning on the ID-Selfie datasets

  1. Set up the dataset paths and the pre-trained model path in config/finetune.py

    Xem thêm: vô hạn tuần hoàn là gì

    # Training dataset path train_dataset_path = '/path/to/training/dataset/folder' # Testing dataset path test_dataset_path = '/path/to/testing/dataset/folder' ... # The model thư mục from which lớn retore the parameters restore_model = '/path/to/the/pretrained/model/folder'

  2. Tune the parameters of loss functions according lớn your dataset in config/finetune.py, e.g.

    # Loss functions and their parameters. losses = { 'diam': {'scale': 'auto', 'm': 5.0, 'alpha':1.0} }

    In our experiments, we found that there is no necessity lớn manually choose "scale". But in some cases one may find it helpful lớn change the "scale" lớn a fixed value. A smaller "alpha" should be favored when the average number of samples per class is larger.

  3. Run the following command in the terminal lớn start fine-tuning:

    python src/train_sibling.py config/finetune.py

Part 3: Feature Extraction

Note: In this part, we assume you are in the directory $DOCFACE_ROOT/

To extract features using a pre-trained model (either base network or sibling network), prepare a .txt tệp tin of image list. The images should be aligned in the same way as the training dataset. Then lập cập the following command in terminal:

python src/extract_features.py \ --model_dir /path/to/pretrained/model/dir \ --image_list /path/to/imagelist.txt \ --output /path/to/output.npy

Notice that the images in the image list follow the same naming convention of the training dataset. That is, ID images should start with "A**" and selfie images should start with "B**".


  • BaseModel (unconstraind face matching): Google Drive | Baidu Yun

  • Fine-tuned DocFace model: (contact the author)


  • Using our pre-trained base model, one should be able lớn achieve 99.67% on the standard LFW verification protocol and 99.60% on the BLUFR protocol. Similar results should be achieved by using our code lớn train the Face-ResNet on Ms-Celeb-1M.

  • Using the proposed Max-margin Pairwise Score loss and sibling network, DocFace acheives a significant improvement compared with Base Model on our private ID-Selfie dataset after transfer learning:

  • Results of DIAM-Softmax and DocFace+ on a combination of ID-Selfie-A, ID-Selifie-B and another larger dataset, most of whose classes have only two images (a pair of ID and selfie):


Yichun Shi: shiyichu at msu dot edu

Xem thêm: cách gấp chim hạc