Reproduce “UncertaintyTrack: Exploiting Detection and Localization Uncertainty in Multi-Object Tracking” on OpenResearch

About this paper

Chang Won Lee, Steven L. Waslander

Researchers at the University of Toronto developed UncertaintyTrack, a collection of extensions that integrates explicit localization uncertainty estimates from a probabilistic object detector into existing multi-object tracking (MOT) frameworks. This approach reduced identity switches by up to 55.1% and improved mean Multi-Object Tracking Accuracy (mMOTA) by 2-3% on challenging datasets like BDD100K and MOT17.

Key insights

• Explicitly modeling localization uncertainty provides meaningful insights into tracking failures, showing that high uncertainty correlates with errors like false positives, false negatives, and ID switches.
• The quality and informativeness of uncertainty estimates from the probabilistic object detector are crucial; UncertaintyTrack's efficacy is dependent on the underlying detector providing accurate and confident uncertainty.
• Bounding Box Relaxation, which expands bounding boxes based on corner uncertainty, is the most impactful component for reducing ID switches and improving overall tracking performance.

Summary by alphaXiv.