Cell tracking via Structured Prediction and Learning

Abstract

In this work, we propose a new joint detection and tracking method for cell tracking. First, we develop a new procedure for generating an over-complete set of detection hypotheses via ellipse fitting, and then, we define several local events and their corresponding labeling variables to account for both biological behavior of cells and segmentation errors. The task of cell tracking is formulated as an integer linear programming problem with constraints and solved efficiently using commercial software. In addition, instead of learning local classifiers independently, we exploit block-coordinate Frank---Wolfe algorithm to learn the optimal parameters of our model under the framework of structured SVM. We also present the kernelized version of the learning algorithm which can boost the tracking performance further. Experimental results on public datasets show that our method is competitive with the state-of-the-art ones.