Learning-Based Shadow Detection in Aerial Imagery Using Automatic Training Supervision from 3D Point Clouds

Shadows, motion parallax, and occlusions pose significant challenges to vision tasks in wide area motion imagery (WAMI) including object identification and tracking. Although there are many successful shadow detection approaches that work well in indoor scenes, close range outdoor scenes, and spaceborne satellite images, the methods tend to fail in intermediate altitude aerial WAMI. We propose an automatic shadow mask estimation approach using self-supervised learning without manual labeling to provide a large amount of training data for deep learning-based aerial shadow extraction. Analytical ground-truth shadow masks are generated using 3D point clouds combined with known solar angles. FSDNet, a deep network for shadow detection, is evaluated on aerial imagery. Preliminary results indicate that training using automated shadow mask self-supervision improves performance, and opens the door for developing new deep architectures for shadow detection and enhancement in WAMI.