PI: Michael Felsberg, Linköping University
co-PI: Anders Heyden, Lund University
Learning geometric representations is a project that aims to solve geometric computer vision tasks that occur in many application domains, ranging from autonomous driving to image-based bio-diversity research. In contrast to previous projects that used geometric constraints in data-driven methods, the proposed project makes use of multi-modal input data over time, e.g. sequences of images and point clouds, that imply a manifold learning task for implicit neural representations. Three research questions need to be addressed within two PhD projects to validate this approach: a) should geometry and appearance be treated separately or jointly, b) how to inject geometric constraints into implicit representations, and c) how can the manifold induced by multi-modality be exploited to extrapolate beyond previously seen data? The project team is based on a well-established collaboration between the mathematical vision group in Lund and the computer vision laboratory in Linköping.
Project number: F3