Fourier-Based Instance Selective Whitening for Domain Generalized Lane Detection

Lane detection represents a fundamental task within autonomous driving. While deep learning has made remarkable advancements in the source domain, its ability to generalize to unseen target domains still poses a challenge. To address this issue, we present a Fourier-based instance selective whitening framework. This framework utilizes the distinct frequencies within the Fourier spectrum to decompose data style into environment and texture styles. Our method preserves semantic features by stabilizing the phase component, while also extending the style through perturbing and amalgamating the amplitude component. Further, we propose a standardized instance selective whitening strategy to analyze overall distributional changes, emphasizing general features and reducing domain-specific information. Our approach is validated through extensive experiments across multiple challenging datasets, such as Tusimple, CULane, and LLAMAS, which demonstrates significant effectiveness when compared to existing methods.