All-Day Multi-Camera Multi-Target Tracking

Paper

This repository is an official implementation of AMCMT model and M3Track dataset.

Abstract

In this paper, we propose ADMCMT, which is the first MCMT tracking model for low-light environments. Specifically, we propose an All-Day Mamba Fusion (ADMF) module to adaptively fuse information from different modalities. Within ADMF, the Lighting Guidance Model (LGM) extracts lighting relevant information to guide the fusion process. Furthermore, the Nearby Target Collection (NTC) strategy is designed to enhance tracking accuracy by leveraging information derived from surrounding objects of target. Experiments conducted on M3Track demonstrate that ADMCMT exhibits strong generalization across different lighting conditions. We also constructed the first Multi-modality (RGBT) Multi-camera Multi-target tracking dataset named M3Track. In addition to incorporating multiple modalities, it also significantly surpasses existing multi-camera multi-object datasets in both scale and diversity.

Main Results

M3Track

Method	MOTA↑	HOTA↑	IDF1↑	DetA↑	MOTP↑	MDA↑	CVIDF1↑	CVMA↑
OSNet	68.03	50.33	61.60	55.71	75.78	0.2756	48.50	28.60
CrossMOT	65.30	44.84	59.91	48.62	75.54	0.2019	51.21	38.51
CT	68.21	50.38	61.17	55.89	75.80	0.4340	53.65	44.71
MvMHAT	66.39	48.60	58.96	54.99	75.77	0.2178	46.64	26.72
AGW	67.62	50.44	62.02	55.54	71.54	0.4235	54.81	45.02
Ours	70.94	56.62	71.47	57.89	77.21	0.6169	68.77	61.81

Plan for Open-Sourcing

Since our work is based on the unpublished GMT, we are unable to release the code now. We will release the code as soon as GMT is published.

M3Track dataset

Introduction

M3Track is the first multi-camera multi-target tracking dataset that includes the infrared modality. Compared with existing datasets, M3Track is larger in scale and contains data captured under different weather conditions, at various times of the day, with varying target densities, and across diverse scenes, offering rich diversity.

Download link

For commercial use, please contact the authors for authorization.

OneDrive: Link

BaiduNetdisk: link    Code: mcmt

Citation

If you find this project useful for your research, please use the following BibTeX entry.

@InProceedings{Fan_2025_CVPR,
    author    = {Fan, Huijie and Qiao, Yu and Zhen, Yihao and Zhao, Tinghui and Fan, Baojie and Wang, Qiang},
    title     = {All-Day Multi-Camera Multi-Target Tracking},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {16892-16901}
}