TrackOpt – Learning to Optimize Physically Constrained Sparse-to-Dense Point Tracking – funded by the BMBF (Federal Ministry of Education and Research)

The tracking of individually occurring points or dense point clouds is highly relevant in a variety of physical and biological problems. Such problems pose a variety of challenges for machine learning methods, since learning must be possible on the basis of often relatively small amounts of annotated training data that is often subject to measurement uncertainties. The aim of this project is to create an efficient method framework that enables new tracking problems to be flexibly integrated and reliably solved. Model-driven, potentially discrete optimization problems are integrated with deep neural networks so that the strengths of both approaches can be exploited. In particular, the integration of model-driven knowledge means that the planned framework can work in a very data-efficient manner. At the same time, learned networks are optimized to predict consistent solutions despite uncertainty and measurement inaccuracy.