Comparison of Semantic Segmentation Approaches for Horizon/Sky Line Detection

TL;DR

This paper compares four autonomous horizon/sky line detection methods, including specialized and general segmentation approaches, on a large diverse dataset to evaluate their accuracy and pixel error for improving geo-localization.

Contribution

It provides a comprehensive quantitative comparison of four recent horizon detection methods, including transfer learning for sky segmentation, on an extensive dataset.

Findings

All methods tested show varying accuracy and pixel error levels.

Transfer learning improves sky segmentation performance.

Specialized horizon detection methods outperform general segmentation in accuracy.

Abstract

Horizon or skyline detection plays a vital role towards mountainous visual geo-localization, however most of the recently proposed visual geo-localization approaches rely on \textbf{user-in-the-loop} skyline detection methods. Detecting such a segmenting boundary fully autonomously would definitely be a step forward for these localization approaches. This paper provides a quantitative comparison of four such methods for autonomous horizon/sky line detection on an extensive data set. Specifically, we provide the comparison between four recently proposed segmentation methods; one explicitly targeting the problem of horizon detection\cite{Ahmad15}, second focused on visual geo-localization but relying on accurate detection of skyline \cite{Saurer16} and other two proposed for general semantic segmentation -- Fully Convolutional Networks (FCN) \cite{Long15} and SegNet\cite{Badrinarayanan15}. Each of the first two methods is trained on a common training set \cite{Baatz12} comprised of about 200 images while models for the third and fourth method are fine tuned for sky segmentation problem through transfer learning using the same data set. Each of the method is tested on an extensive test set (about 3K images) covering various challenging geographical, weather, illumination and seasonal conditions. We report average accuracy and average absolute pixel error for each of the presented formulation.