The article presents a new framework for single-source domain generalization in lane detection. This is important for autonomous vehicles to navigate and localize their position on the road. The framework decomposes data into lane structures and surroundings, enhancing diversity with High-Definition (HD) maps and generative models. Instead of expanding data volume, a core subset of data is strategically selected to maximize diversity and optimize performance. The study shows that this approach improves the generalization performance of lane detection, making it comparable to the domain adaptation-based method.
Publication date: 29 Nov 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2311.16589