The quality of the region segmentation will aect the performance of our shadow
removal as we operate at the region level. Ideally, we want to segment the image
into superpixels that correspond to consistently illuminated regions. That is,
either all pixels in a region are in shadow or all are not in shadow. Furthermore,
we would like to obtain homogeneous regions in terms of material.
To segment the images into regions we use the segmentation method proposed
by Yago [4], where they segment images into regions for shadow detection. The
method consists of a two step process that is fast and robust to the choice of
thresholds. First, SLIC [11] superpixel segmentation is applied to oversegment
the image, obtaining an initial set of superpixels. Then, mean-shift clustering[12]
over the superpixels' mean color in CIELAB color space is performed. Lastly,