SkyCloud: Neural Network-Based Sky and Cloud Segmentation from Natural Images

Abstract

The comprehensive understanding of outdoor scenes is a necessary requirement for a wide variety of applications. For example, semantic segmentation enables applications such as outdoor robot navigation, image stylization, weather fore-casting, or climate monitoring. However, existing outdoor scene understanding models are often less reliable in challenging situations such as changing weather conditions or low light. Additionally, current approaches mainly focus on sky and ground separation and do not incorporate valuable information provided by weather conditions and cloud coverage. To overcome these challenges, we present SkyCloudNet, a multitask neural network architecture that extracts high-level attributes from the input image and utilizes them to improve the robustness of the network to environmental influences. Furthermore, it allows for the segmentation of cloud segments in natural outdoor images. While existing cloud segmentation approaches are limited to cropped sky-only images, our model enables the segmentation from entire landscape images with arbitrary resolution. Next to that, SkyCloudNet achieves state-of-the-art performance in environmental attribute estimation and sky segmentation. As cloud segmentation from natural images has not been addressed in previous literature, we also release the SkyCloud data set consisting of 350 high-resolution outdoor images with dense labels of sky and cloud segments.