Kluger Florian ; Ackermann Hanno ; Yang Michael Ying ; Rosenhahn Bodo

Abstract

The horizon line is an important geometric feature for many image processingand scene understanding tasks in computer vision. For instance, in navigationof autonomous vehicles or driver assistance, it can be used to improve 3Dreconstruction as well as for semantic interpretation of dynamic environments.While both algorithms and datasets exist for single images, the problem ofhorizon line estimation from video sequences has not gained attention. In thispaper, we show how convolutional neural networks are able to utilise thetemporal consistency imposed by video sequences in order to increase theaccuracy and reduce the variance of horizon line estimates. A novel CNNarchitecture with an improved residual convolutional LSTM is presented fortemporally consistent horizon line estimation. We propose an adaptive lossfunction that ensures stable training as well as accurate results. Furthermore,we introduce an extension of the KITTI dataset which contains precise horizonline labels for 43699 images across 72 video sequences. A comprehensiveevaluation shows that the proposed approach consistently achieves superiorperformance compared with existing methods.