UIT- VisDrone- Flood: A Synthesized Aerial Vehicle Detection Dataset Under Flood Conditions

Object detection is a critical task in computer vision, where the dataset's quality is paramount to the model's success. Incorrect or inadequate annotations can significantly impair model performance. In transportation, datasets are typically abundant and diverse, but they predominantly capture normal conditions. Datasets become exceedingly scarce for specific scenarios like floods, especially those with pre-annotated data. Creating a dataset for traffic vehicles in flood conditions requires substantial time, effort, and cost. To address this challenge, we propose a straightforward approach to synthesize a dataset of traffic vehicles in flood conditions. Our method leverages the VisDrone dataset, which contains images of traffic vehicles under normal conditions, and combines it with two advanced technologies: ClimateGAN and the Segment Anything Model (SAM). ClimateGAN accurately and realistically transforms vehicle images from normal to flood conditions, while SAM ensures proper segmentation and annotation of objects within the images. This approach saves time and effort while producing a comprehensive and reliable dataset for research and applications in flood-prone traffic conditions. The UIT-VisDrone-Flood dataset, evaluated using YOLOv10 models, achieved a Mean Average Precision (mAP) of only 28.8, indicating that these models struggle to accurately detect and classify objects in the complex flood conditions represented in this dataset.