Github Repo

Partnership with Grace Benner and Chenchen Lin

UnShade is a naive, yet surprisingly powerful, unsupervised baseline for portrait shadow removal. Using only classical computer vision techniques, such as Poisson blending, PatchMatch, and Gaussian mixture models, UnShade is a fully-unsupervised approach that can outperform unsupervised models in terms of structural similarity and can beat supervised models in terms of perceptual similarity.