One of the major challenges in camera pose estimation and 3D localization is identifying features that are approximately invariant across seasons and in different weather and lighting conditions. In this paper, we present a method for performing accurate and robust six degrees-of-freedom camera pose estimation based only on the pixelwise semantic labelling of a single query image. Localization is performed using a sparse 3D model consisting of semantically labelled points and curves, and an error function based on how well these project onto corresponding curves in the query image is developed. The method is evaluated on the recently released Oxford Robotcar dataset, showing that by minimizing this error function, the pose can be recovered with decimeter accuracy in many cases.