deep-landmark

Predict facial landmarks with Deep CNNs powered by Caffe.

This project is a reimplementation of the paper Deep Convolutional Network Cascade for Facial Point Detection.

Data

All training data can be downloaded from here.

Download the images and extract to dataset with train and test.

modify level1.py, level2.py, level3.py under dataset to change to training datasets.

Train

./bootstrap.sh

This will first generate prototxt files for caffe models and convert training data(images and landmarks) into h5 files. Then We will train the level-1 CNNs and use the result to generate training data for level-2. And for level-2 and level-3 goes the same way.

I strongly suggest you to train every CNN seperately. It's every important to view the loss at first to see if it is stable, if not, stop the training and restart.

View Trainging Logs

I have modified Caffe source code to log the test loss over every test, and I write view_loss.py to plot the loss, all log file are under log so as plot. If the loss plot is unusual, retraining the CNN model is needed.

Caffe will log all stuffs during the network training, you can find the log file under /tmp or you can give Caffe a hit where to save the log file. If you want to see the training loss curve, you should write a program to parse the log file yourself.

Models

All model files are under model, we can modify *.template file to change the caffe network structure for every level.

Results

I have created a web page to test the project, all code are under webapp.

error of every landmark in Level-3

some test

video test

https://youtu.be/oNiAtu0erEk

References

1. Caffe
2. Deep Convolutional Network Cascade for Facial Point Detection