# MO-YOLO: End-to-End Multiple-Object Tracking Method with YOLO and Decoder

**Authors:** Liao Pan, Yang Feng, Wu Di, Liu Bo, Zhang Xingle  
**arXiv:** [2310.17170](https://arxiv.org/abs/2310.17170)  
**Date:** 2023-10-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.17170v3/

## Key points

- Recent advancements in Transformer-based end-to-end models in multi-object tracking (MOT) have shown remarkable performance on challenging datasets like DanceTracker.
- MO-YOLO, an efficient and computationally frugal end-to-end MOT model, was introduced by a team of researchers including Liao Pan, Yang Feng, Wu Di, Liu Bo, and Zhang Xingle.
- MO-YOLO combines principles from GPT, You Only Look Once (YOLO), and RT-DETR while adopting a decoder-only approach for improved efficiency.
- By leveraging the decoder architecture from RT-DETR and key components from YOLOv8, MO-YOLO achieves impressive speed and proficient MOT performance.
- On the Dancetrack dataset, MO-YOLO surpasses existing models like MOTR by achieving over twice the frames per second (MOTR 9.5 FPS vs. MO-YOLO 19.6 FPS).
- MO-YOLO demonstrates significantly reduced training times and lower hardware requirements compared to its counterparts.
- This research presents a promising paradigm for efficient end-to-end MOT systems that prioritize enhanced performance while maintaining resource efficiency.

## Summary

In the field of multi-object tracking (MOT), recent advancements in Transformer-based end-to-end models have shown remarkable performance on challenging datasets such as DanceTracker. However, the high computational demands of these models pose significant challenges in terms of training and deployment. To address this issue, a team of researchers including Liao Pan, Yang Feng, Wu Di, Liu Bo, and Zhang Xingle have introduced MO-YOLO, an efficient and computationally frugal end-to-end MOT model. Drawing inspiration from successful models like GPT, MO-YOLO combines principles from You Only Look Once (YOLO) and RT-DETR while adopting a decoder-only approach. By leveraging the decoder architecture from RT-DETR and incorporating key components from YOLOv8, MO-YOLO achieves impressive speed and proficient MOT performance. In fact, on the Dancetrack dataset, MO-YOLO not only matches but surpasses the performance of existing models like MOTR by achieving over twice the frames per second (MOTR 9.5 FPS vs. MO-YOLO 19.6 FPS). Additionally, MO-YOLO demonstrates significantly reduced training times and lower hardware requirements compared to its counterparts. This research presents a promising paradigm for efficient end-to-end MOT systems that prioritize enhanced performance while maintaining resource efficiency. The innovative approach taken by the authors showcases the potential for future developments in this field to overcome computational limitations and improve overall tracking capabilities in various applications.