Eye Centre Localisation with Convolutional Neural Network Based Regression

This paper introduces convolutional neural network regression models based on the Inception-v3 and the DenseNet architectures for accurate and real-time eye centre localisation. At a normalised error of e < 0.05, the proposed method yields an accuracy of 98.55% on the BioID dataset in a five-fold cross validation test, and 98.50% on the GI4E dataset in a cross-dataset validation test, outperforming the state-of-the-art methods. Both models, capable of running at 44 frames per second, demonstrate an excellent real-time performance. Not only is the proposed method highly accurate and efficient, it does not require invasive and expensive hardware, offering the potential for spawning applications in a wide variety of domains.