Abstract
Unsupervised Domain Adaptation demonstrates great potential to mitigatedomain shifts by transferring models from labeled source domains to unlabeledtarget domains. While Unsupervised Domain Adaptation has been applied to a widevariety of complex vision tasks, only few works focus on lane detection forautonomous driving. This can be attributed to the lack of publicly availabledatasets. To facilitate research in these directions, we propose CARLANE, a3-way sim-to-real domain adaptation benchmark for 2D lane detection. CARLANEencompasses the single-target datasets MoLane and TuLane and the multi-targetdataset MuLane. These datasets are built from three different domains, whichcover diverse scenes and contain a total of 163K unique images, 118K of whichare annotated. In addition we evaluate and report systematic baselines,including our own method, which builds upon Prototypical Cross-domainSelf-supervised Learning. We find that false positive and false negative ratesof the evaluated domain adaptation methods are high compared to those of fullysupervised baselines. This affirms the need for benchmarks such as CARLANE tofurther strengthen research in Unsupervised Domain Adaptation for lanedetection. CARLANE, all evaluated models and the corresponding implementationsare publicly available at https://carlanebenchmark.github.io.