The paper of this Thesis can be found in greek here: https://www.e-ce.uth.gr/wp-content/uploads/formidable/59/Axelos_Christos.pdf

Deep learning and particularly neural networks have offered a new point of view in problem solving. These techniques have been heavily adopted by many applications used in daily life or industry. Gaze estimation belongs to this category of applications. Since neural networks have been used in order to solve problems related to gaze estimation, they continuously provide solutions that outperform the previous ones. This thesis proposes a neural network architecture based on the popular residual networks (ResNets), a novel convolutional network introduced in ILSVRC [1] (2015). Specifically, the proposed method is a ResNet-20 network, which achieves competitive performance compared to the literature. This architecture achieves desirable performance regardless of the environmental conditions (in-the-wild) or the facial characteristics and it can also operate well without any calibration techniques. Finally, this solution can substitute the use of expensive, special hardware when high accuracy is not necessary. As a result, reducing the production cost can make these applications accessible not only to specialized users, but to everyone with a laptop and a web camera.