This research paper presents a new approach to visual multi-object tracking (VMOT), named BASE (Bayesian Approximation Single-hypothesis Estimator). The authors argue that probabilistic tracking methods, which are successful in fields like radar and sonar tracking, have been underused in VMOT. They attribute this to the failure of probabilistic methods to account for specific aspects of VMOT, such as distance in target kinematics and detector confidence. To address these issues, the authors propose BASE, which they claim achieves state-of-the-art results on MOT17 and MOT20 benchmarks. Their code is available on GitHub.
Publication date: 21 Sep 2023
Project Page: https://github.com/ffi-no
Paper: https://arxiv.org/pdf/2309.12035
