CeyMo is a novel benchmark dataset for road marking detection which covers a wide variety of challenging urban, sub-urban and rural road scenarios. The dataset consists of 2887 total images of 1920 × 1080 resolution with 4706 road marking instances belonging to 11 classes. The test set is divided into six categories: normal, crowded, dazzle light, night, rain and shadow.

For more details, please refer to our paper CeyMo: See More on Roads - A Novel Benchmark Dataset for Road Marking Detection.

The train set, the test set and a sample of the CeyMo road marking dataset can be downloaded from the following Google Drive links.

• Train Set - 2099 images (1.17 GB)
• Test Set - 788 images (442 MB)
• Sample - 10 images (5 MB)

A set of raw video clips recorded from the two cameras can be downloaded from here as unlabeled data.

The road marking annotations are provided in three formats: polygons, bounding boxes and pixel-level segmentation masks. The polygon annotations in JSON format are considered as the ground truth and bounding box annotations in XML format and segmentation masks in PNG format are provided as additional annotations. The camera and the vehicle used for capturing each image, and the category (only for test images) are also annotated.

Labelme can be used to visualize the polygon annotations (Images and JSON files should be copied to the same folder) and LabelImg can be used to visualize the bounding box annotations (Images and XML files should be copied to the same folder). The segmentation masks have the following color mapping for the 11 classes in the dataset. Colors are given as RGB color codes.

The column graph (a) shows the frequency of each class in the dataset while the pie chart (b) shows the proportion of each scenario in the test set.

The evaluation script requires the following dependencies to be installed with Python 3.

pip install argparse shapely tabulate

The class-wise, scenario-wise and overall results can be obtained by running the provided python script as follows. <gt_dir> should contain the ground truth json files and <pred_dir> should contain the prediciton json files following the same format as per the ground truth.

python eval.py --gt_dir=<gt_dir> --pred_dir=<pred_dir>

The performance of the four baseline models trained and evaluated on our dataset are as follows.

If you use our dataset in your work, please cite the following paper.

@InProceedings{Jayasinghe_2022_WACV,
    author    = {Jayasinghe, Oshada and Hemachandra, Sahan and Anhettigama, Damith and Kariyawasam, Shenali and Rodrigo, Ranga and Jayasekara, Peshala},
    title     = {CeyMo: See More on Roads - A Novel Benchmark Dataset for Road Marking Detection},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2022},
    pages     = {3104-3113}
}