Foggy License Plates Worldwide: A Comprehensive Dataset
Description
The "Foggy License Plates Worldwide" dataset is a specialized collection designed to advance the recognition and detection of vehicles and license plates under foggy conditions. This dataset includes 4420 2D-RGB images, featuring a diverse array of vehicles from Bangladesh, Thailand, Saudi Arabia, and other regions with English license plates. These images are derived from a secondary dataset and have been augmented using monocular depth estimation to simulate varying degrees of fog, offering a realistic set of data for challenging visibility scenarios. The dataset is not limited to one type of vehicle but includes buses, trucks, CNG vehicles, motorcycles, cars, and standalone license plates, providing a broad spectrum for analysis. While the Bangladeshi subset contains 2754 annotated images, the dataset also includes 388 images from Thailand, 433 from regions with English license plates, and 845 from Saudi Arabia, though the latter does not come with annotations. This variety is crucial for developing robust algorithms that can operate across different regions and vehicle types, especially in foggy conditions. The use of the Monodepth2 Network to artificially introduce fog effects based on depth estimation ensures that the dataset can mimic real-world scenarios, enhancing the development of automated systems for license plate recognition, vehicle detection, and traffic monitoring under adverse weather conditions. By offering a global perspective with plates from multiple countries, this dataset serves as an invaluable resource for researchers and developers in the field of computer vision, aiming to enhance the accuracy and reliability of systems in foggy environments.
Files
Steps to reproduce
To create the "Foggy License Plates Worldwide" dataset, we adopted a systematic approach to gather and augment the data, ensuring reproducibility and transparency in our research process. Here's an outline of how we arrived at our data: 1. Data Collection: The initial dataset, comprising 4420 2D-RGB images of various vehicles and license plates, was sourced from an existing secondary dataset. This dataset includes a variety of vehicles such as buses, trucks, CNG vehicles, motorcycles, and cars, from different countries including Bangladesh, Thailand, Saudi Arabia, and areas with English license plates. 2. Data Augmentation - Monocular Depth Estimation: - We employed the Monodepth2 Network, a model developed by Godard et al., for monocular depth estimation. This technique helps estimate the depth in clear images, which is crucial for simulating foggy conditions. - The depth estimation was used to modify the clarity of the images, introducing varying levels of fog to create a realistic set of data that represents low-visibility environments. This process involved the adjustment of image brightness and contrast based on the depth information to simulate fog conditions accurately. 3. Annotation: - For the Bangladeshi subset of images, we provided detailed annotations to facilitate object detection and recognition tasks. These annotations follow the YOLO (You Only Look Once) format, a popular framework for object detection tasks. - Each annotated file contains information about the location and dimensions of license plates within the images, aiding in the training and testing of machine learning models for license plate detection. 4. Software and Tools: - The Monodepth2 Network was the primary software tool used for depth estimation. This network is known for its effectiveness in depth prediction using a self-supervised learning approach. - For annotation and image processing, standard computer vision libraries and tools were utilized to handle and modify the images according to our requirements. 5. Reproducibility: - To ensure that others can reproduce our research, we maintained a detailed log of all the steps taken, including the parameters used in the Monodepth2 Network and the specifics of the image processing techniques. - The dataset, along with the annotations and a comprehensive description of the methodology, will be made publicly available in a data repository, ensuring that other researchers can access and utilize our dataset for their own research or to validate our findings. By following these steps, we aimed to create a dataset that not only serves the purpose of advancing research in vehicle detection and license plate recognition in foggy conditions but also adheres to the principles of transparency and reproducibility in scientific research.