FaceKeyPoint Detection

A reimplementation of the paper Deep Convolutional Network Cascade for Facial Point Detection.

Data

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

level1.py, level2.py, level3.py under dataset are the codes to gengrate the specific formate *.h5 file for Caffe models.

Train

generate.py, *.template under prototxt are the code and template to generate prototxt files for Caffe models.

./level.pywill train the CNNs and use the mutilprocess to train level-2 and level-3. It will train every CNN seperately.

Test

run.py under test use Caffe model to predict data.

test.py under test use Caffe model to predict keypoints and evaluate the mean error.

show_w&fea.oy under test are the code to visualize the filiters and feature maps.

density_plot.pyunder test are the code to polt the test-results density picture.

Models

All model files generated by Caffe are under model, It contains all the check points intermediate results and the final *.caffe.

Logs

All the results generated on the test images. Including mean-error, Feature map and Filiters Pictures.

References

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