Collecting and labeling training data for vision-based road scene understanding is a major challenge. The most prominent approach is to use manual labeling, though it is clear that scalability of this approach is limited. More scalable alternatives are simulated data and cross-sensor label transfer. In this talk I will present automatically generated ground truth using one or more sensors, primarily dense Lidar. Specifically, I will present the benefits and challenges of this approach for road scene understanding tasks, including general and category-based obstacle detection, free space and curb detection.