The tracking of an unknown number of divisible cells, or of indistinguishable larvae, is an intriguing problem with important applications in developmental biology, ethology and beyond.
I will show some of the best current approaches to solving this problem: firstly, a tracking-by-assignment approach that tries to detect all targets and then links them up while making allowance for undersegmentation, false positive detections and divisions. Secondly, I will present a model that addresses the detection and assignment steps jointly. Both formulations lead to difficult combinatorial optimization problems that can be solved to optimality for modestly sized instances.
Finally, I will demo the former model, which is available as a workflow in the open source "ilastik" package; and will show a possible strategy to identify the less confident parts of a structured prediction for targeted proofreading.