Unsupervised Labeled Lane Markers Using Maps

Abstract

Large and diverse annotated datasets can significantly increase the accuracy of machine learning models. However, human annotations can be cost and time intensive, and generating 3D information and connectivity for image features using manual annotations can be difficult and error-prone. We therefore propose to automatically annotate lane markers in images and assign attributes to each marker such as 3D positions by using map data. Our method projects map lane markers into image space for far distances and relies on a sample-based optimization to refine projections and increase the accuracy of the labels. As part of this work, we publish the Unsupervised LLAMAS dataset of 100,042 labeled lane marker images from about 350 km recorded drives which make this one of the largest high-quality lane marker datasets that is freely available. We estimate that manually annotating a dataset of this size would take several person years. The dataset contains pixel-level annotations of dashed lane markers, 2D and 3D endpoints for each marker, and lane associations to link markers. With the dataset, we create and open source benchmark challenges for binary marker segmentation, lane-dependent pixel-level segmentation, and lane border regression to enable a straightforward comparison of different detection approaches at https://unsupervised-llamas.com.