We present a novel method for single image shadow removal based on region
relighting. We leverage the use of a dataset with annotated shadows to train a
classier that identies non-shadow regions that neighbor shadow regions of the
same material. We propose to use a neighboring lit region to relight a shadow
region. To do so, we rst match the luminance values of the shadow pixels to
the luminance histogram of the lit region. Then, we adjust the shadow region
chromaticities by adding the dierence between the median CIELAB a and b
values of the lit region and the shadow region. However, the image segmenta-
tion often outputs inaccurate boundaries such that shadow(lit) pixels leak into
a lit(shadow) region. Hence, we perform the relighting process only on the core
pixels of the regions. That is, we ignore the outer perimeter pixels of each region.
We iteratively nd pairs of shadow and lit neighbors and relight the shadow re-
gions. Finally, we process the shadow boundaries. Our results outperform the
state of the art in the benchmark dataset[1]. For shadow pixels we obtain a
shadow removal error, measured as Root Mean Square Error (RMSE), of 9.24,
a 21% reduction compared to [10]. This article contains the following main con-
tributions: