An Experimental Evaluation of Different Features and Nodal Costs for Horizon Line Detection

Horizon line detection is a segmentation problem where a boundary between a sky and non-sky region is searched. Conventionally edge detection is performed as the first step followed by dynamic programming to find the shortest path which conforms to the detected horizon line. Recent work has proposed the use of machine learning to reduce the number of non-horizon edges to accurately detect the horizon line. In this paper, we investigate the suitablity of various local texture features and their combinations to reduce the number of false classifications for a recently proposed horizon detection approach. Specifically, we explore SIFT, LBP, HOG and their combinations SIFT-LBP, SIFTHOG, LBP-HOG and SIFT-LBP-HOG as features to train the SVM classifier. We further show that using only edge information as the nodal costs is not enough and propose various nodal costs which can result in enhanced accuracy of the detected horizon line as evidenced by the conducted experiments and results. We compare our proposed formulations with an earlier approach relying only on edges and suffers due to faulty assumptions. We report our comparative results for an image set comprising of mountainous images captured during an outdoor robot exploration of Basalt Hills.