Simple Evaluation of Association Quality in Tracking-by-Detection

Evaluating multiple-object trackers is challenging due to the variable number of quantities involved and the mixed discrete--continuum nature of the problem. Existing methodologies primarily address detection and tracking challenges. These challenges aim at whole computer-vision pipelines as opposed to isolated tracker algorithms. However, modern tracker algorithms have become sufficiently sophisticated to merit a stand-alone analysis. The most critical component of a tracker is the association procedure, as its outcome strongly affects the overall tracking quality. We propose a straightforward quality assessment framework to evaluate the association quality of trackers. The framework relies on a minimal, query-oriented instrumentation of the tracker. This instrumentation exposes the tracker's internal association decisions, allowing for a binary classification of detection-target matches. The proposed methodology is simple to implement, introduces minimal computational overhead, and agrees with the well-known ClearMOT metrics. The comparison is implemented through our accompanying open-source software packages.