Cell Tracking Using Structured Prediction

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 hypothesis via ellipse fitting methods. Then we define several local events and corresponding labeling variables to account for the biological behavior of cells and the imperfection in segmentation, and formulate the task of cell tracing as an integer programming problem with constraints. In addition, instead of learning local classifiers, we exploit a recently proposed block-coordinate Frank-Wolfe algorithm to automatically learn optimal parameters of our model. We also present the kernelized version of the learning algorithm which can boost the tracking performance even further. We conduct extensive experiments on public datasets, showing that our method consistently outperforms traditional countetparts.