Fisheye sky segmentation dataset

This dataset contains 182 fisheye images and their corresponding pixel-level sky-obstacle annotations, designed for precise sky segmentation in hemispherical fisheye imagery. It combines synthetic fisheye images generated from Google Street View panoramas with annotated real fisheye images captured using smartphone fisheye lenses across diverse environments, including rural, semi-urban, and urban scenes. Images are partitioned into training, validation, and test subsets to support model training and evaluation. In addition, the dataset includes pre-trained model weights for multiple U-Net++ EfficientNet backbones (B4-B7) and their associated LightGBM meta-models, enabling direct inference and full reproducibility of the hybrid CNN-LightGBM sky segmentation pipeline.