Benchamark for GRAP-MOT: Unsupervised Graph-based Position Weighted Person Multi-camera Multi-object Tracking in a Highly Congested Space

This dataset is a large-scale collection of precisely annotated images, that can be used both for training and validating person multi-camera tracking algorithms. This benchmark focuses on the problem of multi-camera tracking of people in restricted areas like public transport or indoor locations. Our approach focuses on re-identification based on  heads, thus only the heads of people were labelled.

During making the recordings, we made sure that no people were entering or leaving the area during recording a single take. The dataset consists of videos taken from 3 different angles. Each video is 256 frames long. 

The videos were grouped by the number of people (from 2 to 15), that are in the recorded area.  We have 10 recordings for each group giving in total 140 recordings.

Each image was later manually labelled in the form of bounding boxes. Bounding boxes of the same person in the same take between 3 cameras have the same id on all images. 

Our frame is equipped with 3 cameras. One (Fish Eye) in the middle and two (regular ones) in the corners. The FE camera takes pictures in 1280X960 resolution, while the side ones are in 1920X1080 pixels. All images are stored as .jpg files.

Each image has a corresponding text file stored in labels.tar.gz file. Label files have a following structure. 

Exemplary file:

1      0.455469    0.360938    0.039062     0.065625
0     0.247656    0.476042     0.031250     0.050000
2     0.671875     0.514062     0.051562      0.076042

1st column: number of if. It is unique fir each passenger on a take.

2nd column: X: relative horizontal distance from the left upper corner of an image to the middle of detection

3rd column: Y: relative vertical distance from the left upper corner of an image to the middle of detection

4th column: W  relative width of the detection

5th column: H  relative height of the detection

Unfortunately due to technical limitations could not publish the whole dataset at once. The full dataset can be found under this link "".

In this repository, we provide

• Labelled recordings of people in the restricted area ranging from 2 to 15 on a single take.

• For each number of people, 10 takes were taken

• Each take was taken from 3 overlapping cameras at the same time

Citing our work

and this Zenodo Dataset

@dataset{GRAP_MOT_zenodo,
    author       = {Marek Socha, Michał Marczyk, Aleksander Kempski, Michał Cogiel, Paweł Foszner, Michał Staniszewski},
    title        = {Benchamark for GRAP-MOT: Unsupervised Graph-based Position Weighted Person Multi-camera Multi-object Tracking in a Highly Congested Space},
    month        = DEC,
    year         = 2024,
    publisher    = {Zenodo},
    version      = {1.0.0},
    doi          = {10.5281/zenodo.14526116},
    url          = {https://zenodo.org/records/14526116}
}