Abstract In this seminar, we discuss the most recent developement in the structure from motion research field incorporating unsupervised learning. After a theoretical problem formulation and the definition of an objective function used for estimation error optimization, we present a short summary of the past decade's progress in the training of deep neural networks, emphasizing on its implications for the structure from motion domain in particular. We proceed with a succint overview of the most prominent analytic approaches to depth and ego-motion estimation, as well as their learning-based counterparts. We also provide a comparison of the discussed methods in terms of estimation accuracy, operational robustness and applicability in a variety of environments. We finally delve into a detailed treatment of a newly developed deep network for depth and ego-motion estimation trained solely under self-supervision and present the evaluation results from the original publication.
Nick36/unsupervised-learning-depth-ego-motion
Unsupervised Learning of Depth and Ego-Motion from Monocular Video Using 3D Geometric Constraints (Seminar)