Abstract—Traffic signs are vital in providing important information to drivers, ensuring their safety, and helping them follow the road rules. Object detection algorithms like You Only Look Once (YOLO) are used in autonomous vehicles to monitor traffic sign information. However, most object detection research focuses on identifying traffic signs rather than their physical condition. One major issue with the existing dataset is the lack of data on damaged traffic signs for training, which could adversely affect the performance of the object detection algorithm. To address this problem, our paper comprehensively reviews the Image-to-Image (I2I) algorithm to modify existing traffic sign images to showcase different physical statuses (normal and damaged). We conduct experiments using state-of-the-art unpaired image-to-image translation techniques, UNet Vision Transformer cycle-consistent Generative Adversarial Network (UVCGAN) v2, and Energy-Guided Stochastic Differential Equations (EGSDE) to translate normal and damaged traffic sign images. Our experimental results are evaluated using Fréchet Inception Distance (FID) and side-by-side image comparison. We analyze and discuss possible and future improvements.
 
Keywords—traffic sign detection, image generative, Image-to-Image (I2I), Generative Adversarial Networks (GANs), Cycle Generative Adversarial Network (CycleGAN), diffusion model


Copyright © 2024 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.