Measures for ranking cell trackers without manual validation

Cell tracking is often implemented as cell detection and data association steps. For a particular detection output it is a challenge to automatically select the best association algorithm. We approach this challenge by developing novel measures for ranking the association algorithms according to their performance without the need for a ground truth. We formulate tracking as a binary classi cation task and develop our principal measure (ED-score) based on the de nitions of precision and recall. On a range of real cell videos tested, ED-score has a strong correlation (-0.87) with F-score. However, ED-score does not require a ground truth for computation.