Abstract: In particle analysis, the statistics of various shape variables must be determined, and this requires automatic segmentation of individual particles - instances. The student will be introduced to the problem of instance segmentation using deep neural networks. The goal will be to propose a method that can segment particles of different shapes and spatial configurations; see the motivation picture. The emphasis will be on developing a method that does not require complicated annotation, which is usually necessary for classical instance segmentation. Different neural network outputs will be compared and clustering strategies of the output will be tested.  

Supervisor: Filip Šroubek

Collaborating workplace: Ústav přístrojové techniky AV ČR