Evaluating a GAN for enhancing camera simulation for robotics

• DOI: 10.48550/arxiv.2209.06710
• Publication Date: 2022-01-01
• Authors: Asher Elmquist; Radu Serban; Dan Negruţ
• Keywords: FOS: Computer and information sciences; Computer Science - Robotics; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; Robotics (cs.RO)

Given the versatility of generative adversarial networks (GANs), we seek to understand benefits gained from using an existing GAN enhance simulated images and reduce sim-to-real gap. We conduct analysis in context simulating robot performance image-based perception. Specifically, quantify GAN's ability difference image perception robotics. Using semantic segmentation, analyze training testing, nominal enhanced simulation a city environment. As secondary application, consider use enhancing indoor For this object detection is used enhancement testing. The results presented reduction gap when GAN, illustrate its use.