Monoküler derinlik tahmini için yarı denetimli yinelemeli öğretmen-öğrenci öğrenimi

Advances in robotics area and autonomous vehicles have increased the need for accurate depth measurements. Depth estimation is one of the oldest problems of computer vision area. While the depth can be estimated by using many methods, finding a cheap and efficient way of doing it was studied for many years. Although, depth measurements using Lidar sensors or RGB-D cameras provides accurate results, due to cost and narrow applicability they are not very effective. On the other hand, using deep learning architectures to estimate depth seems to provide a more efficient, cheaper and robust solution compared to other methods. With the progress in deep learning, monocular depth estimation problem has gained a lot of attention. Recently, representation learning methods showed very promising accuracy results in depth estimation from single images. In this thesis, a deep learning based network architecture is proposed for monocular depth estimation problem. Furthermore, the network is trained with an iterative teacher-student learning framework in a semi-supervised manner. To make student networks generalize better than the teacher network, noise is injected during training of student networks. According to evaluation results our proposed model achieves state-of-the-art accuracy in monocular depth estimation.