ICCV 2021 Multi-camera Multiple People Tracking Workshop

Introduction

Multi-camera Multiple People Tracking (MMPTRACK) dataset has around 5 hour videos for training and 1.5 hour videos for validation. The dataset is fully annotated with person bounding boxes and corresponding person id. All videos are recorded with cameras placed in different angle at the scene, with guaranteen that all cameras field of view are connected (one camera has overlapped FoV with at least one of the other cameras). field of view and are well calibrated. All the videos are collected inside Microsoft indoor labs where we built 5 simulated environments, retail, lobby, industry, cafe and office. We asked total 28 persons to participate in recording (14 in training, 7 in validation and 7 in testing). All people were paid and signed agreement to release their data to public for research usage. All participants are in various age, gender and race groups.

We also provide the ID annotations cross the 5 simulated environments of MMPTRACK for retrieval evaluation, denoted as MMP-Retrieval.

Description of available files

Currently we provide:

Synchronized frames extracted from videos with 15FPS, 640x320 resolution. The filename is formatted as rgb_[frameid]_[cameraid].jpg
Calibration files with Pinhole mode. We provide both extrinsic and intrinsic parameters for each camera, as well as space size of the environment and voxel size we used to discretesize the space.

Two types of groud-truth annotations, bounding box annotations for each camera view (rgb_[frameid]_[cameraid].json) and annotations for person footpoint in discretesize coordinate in world coordinates system (topdown_[frame_id].csv).

                                    # bounding box annotations for camera view, json format
                                    {
                                        "4": [478, 40, 630, 206]
                                        ...
                                        'id': [l, t, r, b],
                                        ...
                                    }

                                    # person footpoint annotations, csv format
                                    4,	253,	139,	904.15
                                    ...
                                    id, x, y, z(ignore)
                                    ...

Visualization code and calibration file usage can be found here.

Dataset Download

The datasets can only be used for research purpose. The recipient of the datasets must agree to terms and conditions.

If you are interested in downloading the MMPTRACK dataset, please download and sign the dataset usage Terms and Conditions and email back to us at iccv2021mmp@outlook.com.
Any personal data submitted with the form is subject to the Microsoft Privacy Statement.
If you need the MMP-Retrieval labels used in paper [1], please specify this request in your email.
[1] Huang, Zhipeng, et al. "Lifelong Unsupervised Domain Adaptive Person Re-identification with Coordinated Anti-forgetting and Adaptation." In CVPR, 2022.

Sample Videos Visualization

Dataset Citation

Please cite the following paper when using our benchmark.

MMPTRACK: Large-scale Densely Annotated Multi-camera Multiple People Tracking Benchmark. arXiv:2111.15157

    @misc{han2021mmptrack,
        title={MMPTRACK: Large-scale Densely Annotated Multi-camera Multiple People Tracking Benchmark}, 
        author={Xiaotian Han and Quanzeng You and Chunyu Wang and Zhizheng Zhang and Peng Chu and Houdong Hu and Jiang Wang and Zicheng Liu},
        year={2021},
        eprint={2111.15157},
        archivePrefix={arXiv},
        primaryClass={cs.CV}
    }