The single eye gaze estimation was appear in CVPR 2015 paper author Dr. Xucong Zhang. In 2015, LeNet liked model (6.3 error). In 2017, using VGG16 replaced and impoving to 5.4 error.
Above method was based single eye gaze only. Ignore the information from anther eye. In 2018, a new method that two eyes based which called AR-Net get 5.0 error in MPIIGaze dataset.
In ECCV 2018, a new semantic information based method get 4.56 error.
The accuracy of the above methods mostly between 4 and 5 degrees and seems to be difficult to improve in the further.
Materials:
Method:
Setting the acrylice plane in the front of person, open laser pointer and point random position on the plane. Person watch that laser point and camera capture then close laser and capture again.