CNN based head pose estimator

About

This is a python/mxnet implementation of a very simple CNN that predicts the head pose

num of params < 100K

CNN model:

conv1 3*3*32,(2,2),relu

conv2 3*3*32,(2,2),relu

conv3 3*3*64,(2,2),relu

conv4 3*3*64,(2,2),relu

fc1   128,relu

fc2   2,tanh

dataset:

1.http://www-prima.inrialpes.fr/perso/Gourier/Faces/HPDatabase.html

2.[Biwi Kinect Head Pose Database](http://data.vision.ee.ethz.ch/cvl/gfanelli/kinect_head_pose_db.tgz)

Requirement

-opencv

only tested on 2.4.9.1

-mxnet

only tested on 0.7.0

-mtcnn

I use https://github.com/pangyupo/mxnet_mtcnn_face_detection to do face cropping and alignment

padding = 0.27,desired_size = 64

Test

run:

python main.py

examples from validation set:(green as label,red as prediction)

Notice

so sorry for that my model is not robust in real scene, it seems to be sensitive to the background and lights because of the oversimplified public dataset,maybe we can synthesis more data with better background based on HPDatabase.