An Uncertainty-Aware Performance Measure for Multi-Object Tracking

Evaluating the performance of multi-object tracking (MOT) methods is not straightforward, and existing measures fail to consider all available uncertainty information in MOT context. This can lead practitioners select models which produce estimates lower quality, negatively impacting any downstream systems that rely on them. Additionally, most have hyperparameters, makes comparisons different trackers less straightforward. We propose use negative log-likelihood (NLL) posterior given set ground-truth objects as a measure. measure takes into account sound mathematical manner without hyperparameters. provide efficient algorithms for approximating computation NLL several common algorithms, show some cases it decomposes approximates widely-used GOSPA metric, illustrative examples highlighting advantages comparison other measures.