| dc.description.abstract | Abstract: Safe navigation in Arctic seas requires qualified judgement of the ice conditions of ice around the vessel, a skill that takes years of training to master. At the same time, melting of the polar ice is allowing for more traffic, leading to less experienced operators navigating in the area. To ensure safe passage, the development of new systems for aiding ship operators in the Arctic is needed. This work is a step towards computer-vision assisted navigation for the Arctic, by benchmarking the ability of human experts and novices on the task of classifying ice-objects in images and comparing them to an off-the-shelf computer vision model. This task can be seen as a lower bound on the difficulty of ice object recognition (compared to, say, instance segmentation), and gives an idea of how well computer vision fares in the Arctic in the best case. Our method gives a good indication of the ability of both humans and computers trained on clean images to generalize to the difficult conditions often found in the Arctic. By applying realistic distortions, such as synthetic fog, to the images at test time, we can assess the ability of the human participants to generalize. These results are then compared to measurements from the computer vision system, to evaluate the models' ability to work in previously unseen, difficult conditions. The main contribution of the work is to give a realistic benchmark of how well current computer vision models perform when classifying ice objects compared to humans. Furthermore, we discover which types of images and distortions pose challenges to both humans and computers. Together, these form a knowledge base that can help create new and improved computer vision models tailored for sea ice imaging, towards the goal of computer-aided navigation. | en_US |
