Real-time Recognition of Shadow from Deep Edge Detection

Abstract

In this work we address the problem of fast shadow detection from single images of natural scenes. Different from traditional methods that employ expensive optimization methods, we propose a fast semantic-aware Convolutional Neural Network learning framework which trains on different kinds of patches, while integrating semantic shadow information. We primarily cluster pixels based on their material similarities, then in addition to considering individual regions separately, we exploit a higher level interactions between the neighbouring regions. We process the shadow edge pixels between the segments, and relate the regions together.