Uncertainty Quantification of Discrete Association Problems in Image Sequence-based Tracking
Applications, ranging from tracking molecular motion within cells to analyzing complex animal foraging behavior, require algorithms for associating a collection of spot-like particles in one image with particles contained in another image. These associations are often made via network flow algorithms. However, it is often the case that many candidate association solutions (the output of network flow algorithms) have nearly optimal scores; in this case, the optimal assignment solution is of dubious quality. Algorithms for reliably computing the uncertainty of candidate association solutions are under-developed in situations where many particles are tracked over multiple frames of data. This is due in part to the fact that exact uncertainty quantification (UQ) in large association problems is computationally intractable because the exact computation exhibits exponential dependence on the number of particles tracked. We introduce a technique that can accurately and efficiently quantify association ambiguity (i.e., UQ for discrete association problems) without requiring the evaluation of the cost of each feasible association solution. Our method can readily be wrapped around existing tracking algorithms and can efficiently handle a variety of 2D association problems. The applications presented are focused on tracking molecules in live cells. Our method is validated via both simulations and experiments. The experimental applications aim to accurately form tracks and quantify diffusion of quantum dot labeled proteins from \emph{in vivo} measurements; here association problems involving cost matrices possessing hundreds to thousands of rows/columns are encountered. For such large-scale problems, we discuss how our approach can efficiently and accurately quantify inherent uncertainty in candidate data associations.