Reproduce “MOTRv2: Bootstrapping End-to-End Multi-Object Tracking by Pretrained Object Detectors” on OpenResearch

About this paper

Yuang Zhang, Tiancai Wang, Xiangyu Zhang

MOTRv2 enhances end-to-end multi-object tracking by integrating a powerful pretrained object detector into the MOTR framework, effectively addressing detection limitations while preserving robust association. The approach achieves state-of-the-art performance across DanceTrack, BDD100K, and MOT17/20 benchmarks, demonstrating improved HOTA and mMOTA scores.

Key insights

• Bootstrapping end-to-end MOT with pretrained detectors successfully resolves the conflict between detection and association, allowing end-to-end models to leverage both superior detection and robust global association.
• Using strong detection priors as dynamic proposal queries significantly simplifies the transformer's task of accurately detecting new objects and allows it to focus on association.
• Propagating four-dimensional bounding box information (center, width, height) as anchors across frames is crucial for accurate instance association and localization, outperforming reference point propagation.

Summary by alphaXiv.