Shifting paradigms in multi-object tracking

Laura Leal-Taixé

Technische Universität München, DE

Abstract

Multi-object tracking (MOT) is a core task in computer vision, crucial for applications like autonomous driving or augmented reality. Even after years of study, it still remains a challenging task since it requires simultaneous reasoning about track initialization, identity, and spatiotemporal trajectories. This problem has been traditionally addressed with the tracking-by-detection paradigm, which proposes to split the problems into two subtasks: detection and data association. I will discuss the shift towards more recent learning-based paradigms, most notably, tracking-by-regression, and the rise of a new paradigm: tracking-by-attention. I will focus on the problems in current MOT algorithms and the need to focus on merging the detection and the data association tasks.