Distilling Reflection Dynamics for Single-Image Reflection Removal

Abstract

Single-image reflection removal (SIRR) aims to restore the transmitted image given a single image shot through glass or window. Existing methods rely mainly on information extracted from a single image along with some pre-defined priors, and fail to give satisfying results on real-world images, due to inherent ambiguity and lack of large and diverse real-world training data. In this paper, instead of reasoning about a single image only, we propose to distill a representation of reflection dynamics from multi-view images (i.e., the motions of reflection and transmission layers over time), and transfer the learned knowledge for the SIRR problem. In particular, we propose a teacher-student framework where the teacher network learns a representation of reflection dynamics by watching a sequence of multi-view images of a scene captured by a moving camera and teaches a student network to remove reflection from a single input image. In addition, we collect a large real-world multi-view reflection image dataset for reflection dynamics knowledge distillation. Extensive experiments show that our model yields state-of-the-art performances.