Comparison of semantic segmentation approaches for horizon/sky line detection

摘要: Horizon or skyline detection plays a vital role towards mountainous visual geo-localization, however most of the recently proposed geo-localization approaches rely on user-in-the-loop methods. Detecting such segmenting boundary fully autonomously would definitely be step forward for these localization approaches. This paper provides quantitative comparison four methods autonomous horizon/sky line an extensive data set. Specifically, we provide between segmentation methods; one explicitly targeting problem horizon detection[2], second focused but relying accurate [15] and other two general semantic — Fully Convolutional Networks (FCN) [21] SegNet[22]. Each first is trained common training set [11] comprised about 200 images while models third fourth method are fine tuned sky through transfer learning using same tested test (about 3K images) covering various challenging geographical, weather, illumination seasonal conditions. We report average accuracy absolute pixel error each presented formulation.