Fusion of edge-less and edge-based approaches for horizon line detection

Horizon line detection requires finding a boundary which segments an image into sky and non-sky regions. It has many applications including visual geo-localization and geo-tagging, robot navigation/localization, and ship detection and port security. Recently, two machine learning based approaches have been proposed for horizon line detection: one relying on edge classification and the other relying on pixel classification. In the edge-based approach, a classifier is used to refine the edge map by removing non-horizon edges. The refined edge map is then used to form a multi-stage graph where dynamic programming is applied to extract the horizon line. In the edge-less approach, classification is used to obtain a confidence of horizon-ness at each pixel location. The horizon line is then extracted by applying dynamic programming on the resultant dense classification map rather than on the edge map. Both approaches have shown to outperform the classical approach where dynamic programming is applied on the non-refined edge map. In this paper, we provide a comparison between the edge-less and edge-based approaches using two challenging data sets. Moreover, we propose fusing the information about the horizon-ness and edge-ness of each pixel. Our experimental results illustrate that the proposed fusion approach outperforms both the edge-based and edge-less approaches.