Deep learning algorithms for single image depth estimation from UAV images
Programming skill is also mandatory (preferably Python), experience with deep learning frameworks is highly preferred (PyTorch, Keras and Tensorflow)
The estimation of depth from a single image have been a Computer Vision topic for decades. Many approaches such as Shape-from-Shading have been proposed to infer from the illumination of the scene the size and distance between objects. This kind of study has clear applications in many domains such as indoor navigation, collision avoidance and safety in general: this has continuously pushed the development of newer and more sophisticated approaches. More recently, the impressive development of Convolutional Neural Networks has shown how deep learning can largely outperform traditional algorithms in many applications such as image segmentation, classification, object detection and tracking, etc. In this regard, an increasing number of papers dealing with Single Image Depth Estimation (SIDE) implementations have been published presenting supervised and self-supervised methods. Although the first results are encouraging, many problems still need to be solved, in particular when using outdoor and airborne data. The aim of this MSc proposal is to investigate the feasibility and accuracy of the SIDE algorithms using UAV in input.
The student will initially revise the existing literature on the SIDE algorithms. The use of already existing code and available datasets to train these algorithms will be the starting point of the work. The UAV datasets will be available on-line and also provided by the supervisors. Both supervised and self-supervised algorithms will be considered in this study in order to define which solution can guarantee the best performance. An additional challenge will be to specifically evaluate the suitability of SIDE for UAV applications. The expected output will be a trained CNN able to estimate the distance of objects from single images.