Horizon line detection using supervised learning and edge cues

Abstract Traditionally, edge detection has been extensively employed as the basic step for the horizon line detection problem. However, generally such methods do not discriminate between edges belonging to horizon boundary and others due to clouds or other natural phenomenon. Additionally, most edge based methods suffer more in the presence of edge gaps. To address these issues, we propose an edge-less horizon line detection approach based on pixel classification, hence not relying on edge information. The key idea is formulating a multi-stage graph using classification maps, instead of edge maps, where each node cost reflects the likelihood of pixel belonging to the horizon boundary. The shortest path is found in the formulated multi-stage graph using dynamic programming which conforms to the detected horizon line. We demonstrate the performance of the proposed approach on two challenging data sets and provide comparisons with two edge-based methods: one relying on edge detection while the other based on edge classification. Overall, the proposed approach achieves comparable performance against carefully crafted edge based formulations. A by-product of the edge-less approach is its capability of associating a confidence level with the found solution, which can be used to confirm the presence or absence of a horizon line in a given image. The method is also capable of dealing with partial horizon line in an image. To further improve the detection performance, we propose a fusion strategy which combines both edge-based and edge-less information. Extensive evaluations, including a publicly available data set, illustrate the superiority of the proposed fusion approach.