GAN-Driven Vehicle Inpainting for Mobile Mapping Scenes

Mobile mapping systems (MMS) capture geospatial data using vehicle-mounted sensors and cameras for generating digital maps. However, the collected images contain sensitive information requiring privacy protection methods. Although object removal approaches have been proposed, these methods may not be well-suited for MMS imagery due to their unique characteristics and processing requirements. In this study, we propose a Generative Adversarial Network (GAN)-based technique tailored to remove unwanted vehicles from MMS scenes. Our approach leverages Pix2Pix, a GAN architecture for image-to-image translation. To handle the challenge of inpainting full panoramic scenes, we extract viewports containing target vehicles from the raw panoramic image. Subsequently, a Pix2Pix-based inpainting network is employed to inpaint the target object in viewport. We enhance the standard Pix2Pix framework by incorporating adversarial feature matching to generate realistic results. Experimental results demonstrated that our method preserves the output quality while efficiently processing large raw MMS image datasets, outperforming traditional methods.