The gaze direction algorithm is not robust in the following ways :
Detects eyes irregularly with low lighting/profile faces. (Can be solved with adding more classifiers with flags, will increase code length. Any other suggestions?)
Many eyes detected over one eye. Creating mutually exclusive blocks can be an option.
Detecting profile eyes. New classifier needs to be developed? Or use one from the pipeline if exists?
The gaze direction algorithm is not robust in the following ways :