This introduction provides a short, qualitative exploration of results and highlights the main findings. You are encouraged to looks through the results yourself and explore the differences between the presented methods and their success/failure cases.
We have gathered data for our experiments by combining existing datasets (e.g. one provided by Guo at al.) with images we have captured ourselves. The main motivation behind adding new images was the relatively small number of images of soft shadows that we are mainly interested in.
We have processed all the data with our method as well as two variations of Guo et al. (the original, where the shadows are detected automatically and themodified one, with manually-specified shadow masks also used in out method). Additionally we have processed a subset of the data with the method of Arbel and Hel-Or.
The cases where our technique fares worse than other methods are less common. They can usually be explained by
Furter, the method of Guo et al. seems to be more robust than that of Arbel and Hel-Or.
One possible reason for poor results in several images unshadowed by the original method of Guo et al. is that their automatic detection sometimes fails (especially on very soft shadows), leading to constraints that can only be satisfied by extreme color changes. A common problem seems to be that majority of the image is detected as in-shadow and therefore the method has no correct unshadowed regions to match the intensity.