Sky detection by effective context inference

Accurate and efficient detection of sky region plays an important role in many vision applications, such as scene parsing, image retrieval, and robot navigation. This paper presents an novel algorithm for sky detection by combining the merits of superpixels and context inference. Different from most existing solutions our method first produces the coarsely segmented sky regions by using a well-designed contextual inference model, and then uses the graph cut to refine the sky region. Moreover, a relatively large sky database is assembled for model training and evaluation of the detection algorithm. Empirically, we conduct extensive experiments on the sky dataset, and the results indicate that our approach has a higher detection accuracy than the state-of-the-art methods under various weather conditions.