# MOTRv2: Bootstrapping End-to-End Multi-Object Tracking by Pretrained Object Detectors

**Authors:** Yuang Zhang, Tiancai Wang, Xiangyu Zhang  
**arXiv:** [2211.09791](https://arxiv.org/abs/2211.09791)  
**Date:** 2022-11-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.09791v1/

## Key points

- Authors propose a pipeline for end-to-end multi-object tracking (MOT) using a pretrained object detector
- Existing end-to-end methods have poor detection performance compared to tracking-by-detection counterparts
- Proposed pipeline, MOTRv2, uses an additional object detector to generate proposals as anchors and adopts the anchor formulation of queries
- This modification improves overall performance of MOTR while maintaining its end-to-end feature
- MOTRv2 achieves state-of-the-art performance on DanceTrack and BDD100K datasets with 73.4% HOTA on DanceTrack ranking first in the 1st Multiple People Tracking in Group Dance Challenge
- Code for implementing MOTRv2 is available at https://github.com/megvii-research/MOTRv2

## Summary

In their paper, "MOTRv2: Bootstrapping End-to-End Multi-Object Tracking by Pretrained Object Detectors," authors Yuang Zhang, Tiancai Wang and Xiangyu Zhang propose a simple yet effective pipeline for end-to-end multi-object tracking (MOT) using a pretrained object detector. Existing end-to-end methods such as MOTR and TrackFormer have been shown to be inferior to tracking-by-detection counterparts due to poor detection performance. The proposed pipeline, MOTRv2, adopts the anchor formulation of queries and uses an additional object detector to generate proposals as anchors. This provides detection prior to MOTR and greatly eases the conflict between joint learning detection and association tasks in MOTR. The authors demonstrate that this simple modification improves the overall performance of MOTR while maintaining its end-to-end feature. MOTRv2 is evaluated on large scale benchmarks including DanceTrack and BDD100K datasets. The pipeline achieves state of the art performance on both datasets with 73.4% HOTA on DanceTrack ranking first in the 1st Multiple People Tracking in Group Dance Challenge. The authors hope that their approach can provide new insights into the end to end MOT community. Code for implementing MOTRv2 is available at https://github.com/megvii-research/MOTRv2. Overall, this paper presents a promising solution for improving end to end multi object tracking using pretrained object detectors.