Estimating intrinsic characteristics of images for shadow removal

Existing works show that shadow removal tasks can benefit from the physical illumination model on the formation of shadows. Inspired by prior works that recover the intrinsic characteristics by decomposing an image for its reflectance and illumination components, we study a variant of shadow illumination model that can better reflect the complexity in the real world – in this model, shadow-free pixels can be expressed by a translation formed of reflectance and illumination components. Based on the new illumination model, we develop a new LR-ShadowNet, which contains two sub-nets for estimating the illumination and reflectance components, respectively, and one sub-net for refining the shadow-removal result. Besides, several mask guidance module are incorporated into LR-ShadowNet for guiding components estimation and result refinement based on shadow region information. The whole network is trained in an end-to-end fashion guided by the shadow masks. Extensive experiments on the ISTD dataset and SBU-Timelapse dataset show that the proposed LR-ShadowNet achieves competitive performance with less computational cost and strong generalisation ability.