GPC: A Generated Dataset for the Detection of Road Potholes and Cracks

Finding potholes and cracks is always crucial for maintaining roads and ensuring traffic safety. Pothole and crack detection become essential with the advancement of driverless technologies. Pothole and crack detection with deep learning is gaining popularity. However, due to the high cost of data collection, there are not many high-quality data sets of road potholes and cracks in the visual field. As a result, this study develops a new GPC dataset based on the diffusion model, which enhances the original real data with text-generated images. Multiple trials are performed using YOLOv3, YOLOv5, Faster R-CNN, Mask R-CNN with Swin Transformer, and SSD and the results are evaluated. mAP without generated data is 1% to 2% lower than GPC datasets.