This repository shows how to build a face tracker, following:

Mnemonic Descent Method: A recurrent process applied for end-to-end face alignment
G. Trigeorgis, P. Snape, M. A. Nicolaou, E. Antonakos, S. Zafeiriou.
Proceedings of IEEE International Conference on Computer Vision & Pattern Recognition (CVPR'16).
Las Vegas, NV, USA, June 2016.

To install most requirements:

pip install -r requirements.txt

To install OpenCV, the recommended option (through anaconda) is:

 conda install -c menpo opencv3=3.1.0 

To install tensorflow:

This face tracker is trained on the which is available for non commercial purposes.

We use the aligned images (crop + alignment of the face)

If you want to train the model from scratch, you will need to download the dataset from the link above.

You should configure the data folder as follows:

├── data
    ├── RawData
        ├── Anno
            ├──list_attr_celeba.txt  
            ├──list_landmarks_align_celeba.txt  
            └── list_landmarks_celeba.txt
        ├── Eval
            └──list_eval_partition.txt
        └── img_align_celeba

Unzip the aligned images in the img_align_celeba folder and you should be good to go !

In src/model:

python make_dataset.py

In src/model:

python train.py

You may need to change some parameters like the batch size depending on your GPU.

In models:

tensorboard --logdir=train

Then in your web browser, go to http://0.0.0.0:6006 to access tensorboard

Assuming you are sitting in front of your webcam:

In src/model:

python tracker.py

You can modify the choice of model by specifying your model checkpoint at the beginning of the script

tf.app.flags.DEFINE_string('pretrained_model_checkpoint_path', '',
                           """If specified, restore this pretrained model """
                           """before beginning any training.""")

By default, it is set to the pre-trained model that you can find in models

MDM are a clever implementation of a descent method that seeks to align a reference shape (i.e. a collection of landmarks on the face) to a given image.

The alignment is predicted by a combination of a CNN (for feature extraction) and an RNN (to map the extracted feature to a prediction of the correct shape). Typically, this is done in iterative steps (predict a new location for the points, extract features from this new location and repeat). The use of an RNN allows us to optimize all steps simultaneously.

The facetracking app then follows the following method:

• Locate the face with OpenCV's face detector
• Crop the video around the face
• Starting from an initial estimation of the shape, predict the location of landmarks
• For subsequent frames, use the previous location as a starting point for the algorithm.

• The tracker is not very robust to rotation as it was trained on aligned images. This could be solved by data augmentation.
• A failure checker is yet to be implemented (i.e. provide a way to restart if the tracker can't converge) so the program may crash under challenging conditions.
• Only 5 landmark points are provided in the CelebA dataset. This does not help robustness either (typically, more than 40 landmark points are used, which provides a much finer-grained sampling of the image, and possibly better predictions).