Supervisors: Dr. Ajad Chhatkuli, Dr. Thomas Probst, Prof. Luc Van Gool
In this work we tackle the problem of line detection and line clustering with deep neural networks, an approach that has already shown great success in computer vision. While being challenging problems, line detection and clustering in images are extremely useful for several 3D vision applications such as vanishing point detection, camera calibration and camera localization. Like points, lines are primitives that are useful for matching, registration and reconstruction. Two key reasons justify our use of deep neural networks for the task: first, previous works such as D-SAC, show that geometric properties and functions can be learned efficiently. Second, certain domain or environment specific patterns can be learned by deep neural networks during training time, giving an edge over purely algebraic approaches. There are further more reasons such as inference speed and differentiability for using neural networks on well-constrained geometric problems. This motivates us towards our goal: developing self-supervised networks for line detection and line clustering. We explore two kinds of approaches for line detection, while for clustering, we separate them into intersecting lines, parallel lines and others in images using a PointNet-based network. With an assumption of the Manhattan world, such a clustering is directly useful for many 3D vision tasks. We show with experiments that both our line detection and clustering performs comparably with state-of-the-art approaches.