Gaze-Tracking based on head orientation and eye orientation
Reference paper: Monocular Free-head 3D Gaze Tracking with Deep Learning and Geometry Constraints
The angle of the gaze can be calculated by an angle of the head gesture and angle of the eye turn. Here, we will crop out the face and eye parts according to the key points of the face, and then send them to the two subnets face_net and eye_net respectively. The two sub-networks output the head and eye orientation angles, expressed by the longitude 'lo' and the latitude 'la'. Then combine these two angles to calculate the final gaze angle 'lo','la'.
train_head&eye.py
Train the head network to predict head orientation; Train the eye network to predict eye orientation. This code contains data augmentation.
cal_gaze.py
Combine head angles and eye angles to calculate the final gaze angle.
direct_train_gaze.py
Directly predict the gaze, without head-net and eye-net. This may cause overfitting.
extract_feat.py
Detect key-points from face image using face-alignment
train_head_with_landmark.py
Train the GBDT(light GBM) model to predict head angle with key-points.
On this data set, the error value of the angle of gaze is within 10 degrees.
If you want to know more details, please read gaze_tracking_ๆฅๅ.pdf (Chinese).