linhmv/VisualMOT

Beyond the Survey: A Systematic Empirical Study of Detection and Association in Visual MOT

Official implementation repository for the paper "Beyond the Survey: A Systematic Empirical Study of Detection and Association in Visual MOT", accepted for publication in Artificial Intelligence Review.

Note: This repository also includes the official Python implementation for:
"Adaptive Confidence Threshold for ByteTrack in Multi-Object Tracking" (ICCAIS 2023, arXiv 2312.01650) at this folder bytetrack.

---

Overview

This repository provides tracking implementations for algorithms evaluated in our study: SORT, DeepSORT, MOTDT, FairMOT, ByteTrack (AdaptByteTrack), OCSORT, DeepOCSORT, and VisualRFS. We provide CMC (Camera Motion Compensation) for ByteTrack (AdaptByteTrack), VisualRFS, and OC-SORT. Evaluation Scores/Metrics are:

• CLEAR, IDF1, HOTA evaluation available at JonathonLuiten/TrackEval.
• Tracking Effort Measure TEM (vpulab/MOT-evaluation).

For other evaluated methods, please clone their respective repositories and follow the authors' original execution instructions.

---

Datasets & Detection Outputs

Pre-extracted Detections can be downloaded from Hugging Face (linhmv/VisualMOT).

We provide extracted detection files with confidence scores $[0, 1]$ in the (./dets/) directory:

Detector	Venue / Source	Paper / Link
POI: detector_poi	ECCV 2016	arXiv:1610.06136
JDE: detector_jde	ECCV 2020	arXiv:1909.12605
TraDeS: detector_trades	CVPR 2021	arXiv:2103.08808
FairMOT: detector_fairmot128	IJCV 2021	arXiv:2004.01888
GSDT: detector_gsdt	ICRA 2021	arXiv:2006.13164
CSTrack: detector_cstrack	TIP 2022	arXiv:2010.12138
YOLOX: detector_bytetrack	ECCV 2022	arXiv:2110.06864
YOLOv11: detectors_yolov11	arXiv 2024	arXiv:2410.17725

• POI: Relies on ETHZ, Caltech Pedestrian, and a self-collected surveillance dataset.
• TraDeS: Utilizes a CrowdHuman pre-trained model for 2D tracking alongside the MOTChallenge dataset.
• JDE, FairMOT, CSTrack, GSDT: Fine-tuned on the "Mix of Six" dataset, which combines Caltech Pedestrian, CityPersons, ETHZ, MOTChallenge, CUHK-SYSU, and PRW.
• YOLOX: Trained on a combination of MOTChallenge, CrowdHuman, CityPersons, and ETHZ.
• YOLOv11: YOLO11x, Only trained on COCO Detection Dataset

Evaluation Datasets

Dataset Name	Year	Source
MOTChallenge	2016, 2017, 2020
DanceTrack	CVPR 2022	DanceTrack/DanceTrack
SportsMOT	ICCV 2023	MCG-NJU/SportsMOT
CrowdTrack	arXiv 2025	loseevaya/CrowdTrack

---

Evaluated Tracking Algorithms

1. Analytical Data Association

Hand-crafted motion & appearance models

Method	Ref. Index	Year	Source Code
SORT	[23]	ICIP 2016	abewley/sort
DeepSORT	[27]	ICIP 2017	nwojke/deep_sort
MOTDT	[28]	ICME 2018	longcw/MOTDT
FairMOT	[29]	IJCV 2021	ifzhang/FairMOT
ByteTrack	[24]	ECCV 2022	FoundationVision/ByteTrack
AdaptByteTrack	[75]	ICCAIS 2023	linh-gist/AdaptConfByteTrack
OCSORT	[25]	CVPR 2023	noahcao/OC_SORT
DeepOCSORT	[30]	ICIP 2023	gerardmaggiolino/deep-oc-sort
StrongSORT	[32]	TMM 2023	dyhBUPT/StrongSORT
VisualRFS	[1]	PR 2024	linh-gist/VisualRFS
HybridSORT	[31]	AAAI 2024	ymzis69/HybridSORT
TrackTrack	[26]	CVPR 2025	kamkyu94/TrackTrack

2. Deep Learning Data Association

Learned feature-based association

Method	Ref. Index	Year	Source Code
SUSHI	[33]	CVPR 2023	dvl-tum/sushi
LTTrack	[34]	TCSVT 2024	linjiaping1/LTTrack
LG-MOT	[35]	TCSVT 2025	weslee88524/lg-mot

3. End-to-End (E2E) Data Association

Joint detection and association learning

Method	Ref. Index	Year	Source Code
MOTR	[17]	ECCV 2022	megvii-research/MOTR
MeMOTR	[41]	ICCV 2023	mcg-nju/memotr
MOTIP	[42]	CVPR 2025	MCG-NJU/MOTIP
CO-MOT	[43]	ICLR 2025	BingfengYan/CO-MOT
SambaMOTR	[39]	ICLR 2025	mattiasegu/sambamotr

Usage

1. Set Up Python Environment

   • Create a conda Python environment and activate it:

     conda create --name virtualenv python==3.8.0
     conda activate virtualenv
     

   • lone this repository recursively to have pybind11

     git clone --recursive https://github.com/linh-gist/AdaptConfByteTrack.git
     

   • Install Packages

     numpy==1.23.1
     opencv-python==4.9.0.80
     loguru==0.7.2
     scipy==1.10.1
     lap==0.5.12
     cython_bbox==0.1.5
     matplotlib==3.5.3
     filterpy==1.4.5
     motmetrics==1.4.0
     openpyxl==3.1.5
     pycocotools==2.0.7
     tabulate==0.9.0
     # git clone https://github.com/JonathonLuiten/TrackEval.git
     # cd TrackEval, python setup.py build develop
     

2. Prepare Data

   • Datasets:
     • MOT16, MOT17, MOT20, DanceTrack, SportsMOT, CrowdTrack
     • You can also run with your custom dataset but need a detector

3. Run the Tracking Demo

   • Change parameters in make_parser() in track.py such as use_gmc, data_dir (MOTChallenge GT data)
   • Run python track.py

Citation

If you find this project useful in your research, please consider citing by:

@article{van2026beyond,
  title={Beyond the Survey: A Systematic Empirical Study of Detection and Association in Visual MOT},
  author={Linh Van Ma and Juhua Hu and Wei Cheng and Unse Fatima and Moongu Jeon},
  booktitle={Artificial Intelligence Review},
  year={2026},
  publisher={Springer}
}
@inproceedings{van2023adaptive,
  title={Adaptive Confidence Threshold for ByteTrack in Multi-Object Tracking},
  author={Linh Van Ma and Muhammad Ishfaq Hussain and JongHyun Park and Jeongbae Kim and Moongu Jeon},
  booktitle={2023 12th International Conference on Control, Automation and Information Sciences (ICCAIS)},
  pages={370--374},
  year={2023},
  organization={IEEE}
}

Acknowledgement

A part of the code is borrowed from SORT, DeepSORT, MOTDT, FairMOT, ByteTrack, OCSORT, DeepOCSORT, and VisualRFS. Thanks for their wonderful works.