Matthias Springstein, a research assistant from the Visual Analytics Research Group at Technische Informationsbibliothek (TIB) – German National Library of Science and Technology, presented the paper “Estimating Relative Depth in Single Images via Rankboost” at the IEEE International Conference on Multimedia & Expo in Hong Kong in July 2017. The contribution is the result of a collaborative study involving researchers such as Professor Hüllermeier from Paderborn University and Professor Dembczyński from Poznań University of Technology (Poland). Previous methods have usually regarded the problem of depth estimation in a single image as a regression problem.
An innovative approach is presented in the paper, where the task is modelled as a ranking problem and the relative depth or arrangement of regions in an image is determined. This approach offers several advantages: for example, it makes it much easier to produce training data for supervised machine learning methods. The method is based on simple monocular depth features, which are motivated by human (monocular) depth perception. Among other things, the system exploits the vanishing points of an image to draw conclusions about the depth of an image region. The resulting individual depth features are deployed together with the features of a deep neural network to calculate the depth map using a ranking method (RankBoost). Experimental results demonstrated that the approach is able to improve performance of state-of-the-art reference methods. The approach can be applied in several ways, such as in automatic scene interpretation.