Convolutional Neural Network Training System For Eye Location On Infrared Driver Recordings Using Automatically Generated Ground Truth Data

One of the most important factors in being able to train neural networks with good accuracy is the quality of training data. This is the reason why in every project with such purpose a very big part of the effort is to label data, double check and clean the labeled data and finally select the datasets which will be used for training of the neural networks. This paper presents promising results of an automatic system which selects and uses the data from a ground truth data generator for eye location on infrared driver recordings to train neural networks. The quality of the training data is based on the quality of the generator and it is not altered in any way by any human factor. All the selection of the training dataset is done automatically and the system outputs different neural networks, some of them with very good accuracy of eye detection. This paper presents an automatic system that reduces the human effort for labeling and selection of datasets and outputs neural networks with good accuracy for eye location on infrared driver recordings.