Workman Scott Zhai Menghua Jacobs Nathan

Abstract

The horizon line is an important contextual attribute for a wide variety ofimage understanding tasks. As such, many methods have been proposed to estimateits location from a single image. These methods typically require the image tocontain specific cues, such as vanishing points, coplanar circles, and regulartextures, thus limiting their real-world applicability. We introduce a large,realistic evaluation dataset, Horizon Lines in the Wild (HLW), containingnatural images with labeled horizon lines. Using this dataset, we investigatethe application of convolutional neural networks for directly estimating thehorizon line, without requiring any explicit geometric constraints or otherspecial cues. An extensive evaluation shows that using our CNNs, either inisolation or in conjunction with a previous geometric approach, we achievestate-of-the-art results on the challenging HLW dataset and two existingbenchmark datasets.