Time-series microscopy data has the potential yield critical insights into the function of biological systems. However, given the large size of the data produced, computational multiple-object tracking (MOT) algorithms are becoming more and more essential. This often requires custom algorithm design or time-consuming parameter tuning for existing methods . Furthermore, MOT algorithms are prone to systematic mistakes related to the type of motion they are designed to capture . These biases can make it difficult to trust computed tracks to assess differences in cell (or organelle) motion between experimental conditions. We therefore needed a method by which to assess accuracy without the onerous (and often intractable) task of producing ground truth tracks.
We present our ongoing work to estimate tracking accuracy by human proofreading, motivated by our need to compare methods for tracking platelet cells in a growing thrombus (blood clot). Our analysis considers incorrect assignments (false positive rate), false termination, and failure to make correct linkages (false negative rate). Our current tool provides a straightforward method for evaluating tracking accuracy without ground truth. Importantly, proofreading with our method generates sparse but reliable ground truth data, which can then be used for automated parameter optimisation and deep learning. Our tool is available as open source Python code on GitHub.
 Ulman, V., et al. 2017. Nature methods, 14(12), pp.1141-1152.
 Chenouard, N., et al. 2014. Nature methods, 11(3), pp.281-289.