{
    "markers": [
        {
            "id": 29,
            "corners": [
                [
                    481.7849807739258,
                    261.83265686035156
                ],
                [
                    516.1624603271484,
                    203.50667572021484
                ],
                [
                    451.4819869995117,
                    169.94456100463867
                ],
                [
                    425.1454734802246,
                    229.83663177490234
                ]
            ],
            "rot": 2
        },
        ...

Detection datasets (such as det_luma.tar.gz) are provided in the format expected for the training of YOLOv8 (by Ultralytics). The directory structure is the following:

det_luma
├── train
│   ├── images
│   │   ├── 000000000089.jpg
│   │   ├── 000000000315.jpg
│   │   ├── 000000000349.jpg
            ...
│   └── labels
│       ├── 000000000089.txt
│       ├── 000000000315.txt
│       ├── 000000000349.txt
            ...
└── valid
    ├── images
    │   ├── 000000000486.jpg
    │   ├── 000000000612.jpg
    │   ├── 000000000913.jpg
            ...
    └── labels
        ├── 000000000486.txt
        ├── 000000000612.txt
        ├── 000000000913.txt

The images directory contains the augmented frames, while the labels directory stores the ground-truth marker locations in the following format: The first value (class_id) is always 0 since we are only considering a single class for this detection task. The following four values are normalized with respect to the image dimensions: x_center and marker_width are normalized by the image width, while y_center and marker_height are normalized by the image height. See the following example for clarification:

0 0.7353968232870102 0.599690580368042 0.1422140419483185 0.25524471071031357
0 0.4113114923238754 0.8842228227191501 0.031858795881271364 0.07755138079325358
0 0.4016574129462242 0.701622458299001 0.04431436359882355 0.08198304971059163
0 0.865299366414547 0.7644710514280532 0.03821879923343659 0.07337598270840115
0 0.2047746479511261 0.7245340400271946 0.12338865995407104 0.21934681998358832
0 0.32464619278907775 0.7738529602686565 0.03873543739318848 0.06867445309956868
0 0.740096740424633 0.21104383733537463 0.03647349178791046 0.05871715015835232
0 0.33322680816054345 0.13155997859107124 0.028741587698459626 0.051002592510647246

Finally, the regression datasets (such as reg_luma_bc.tar.gz) are provided as a collection of crops (64 x 64 images), each one containing a marker from an image of the matching detection dataset (e.g. reg_luma_bc.tar.gz and det_luma_bc.tar.gz). Labels for these crops are provided in CSV files, with separate files for the training and validation sets. The structure of these CSV files is the following: the first column contains the name of the crop, which corresponds to the source image name followed by a two-digit crop index. This is followed by the x and y coordinates for each of the four corners of the marker, the marker’s rotation, and its ID. See the following example for clarification:

pic,c1_x,c1_y,c2_x,c2_y,c3_x,c3_y,c4_x,c4_y,rot,id
000000492129_00.jpg,0.40259319265856375,0.15452799356724128,0.1545299431229553,0.5946699372972077,0.5912928629425661,0.8454720064327587,0.8454700568770447,0.40691390662961124,3,43
000000492129_01.jpg,0.15401030670038238,0.4037121428420323,0.4583366336930896,0.845905562858047,0.8459896932996176,0.560064082549943,0.559349025823742,0.1540944371419531,2,65
000000492129_02.jpg,0.15134224899058682,0.46932990290960436,0.4602298844365988,0.8486538262764869,0.8486577510094132,0.5485276412548922,0.5280271849356097,0.15134617372351275,3,1

Code

To support the research community and encourage exploration, we have provided access to the code used for the creation of the dataset through our GitHub repository.

•  DeepArUco++ GitHub

Download our dataset

The pre-computed version of this dataset used in our work can be downloaded through Zenodo

Read our work

A preprint of our paper can be accessed through ArXiv: link

Citing

If you use this work in your research, you must cite:

1. Rafael Berral-Soler, Rafael Muñoz-Salinas, Rafael Medina-Carnicer, Manuel J. Marín-Jiménez, DeepArUco++: Improved detection of square fiducial markers in challenging lighting conditions, Image and Vision Computing, Volume 152, 2024, 105313, ISSN 0262-8856, https://doi.org/10.1016/j.imavis.2024.105313

Contact