Supplementary material for Shapecollage : occlusion-aware , example-based shape interpretation

As the paper explains, the basic idea is to define the local appearance score as the average difference in pixel value between the test image and the candidate patch, for pixels where ownership is nonzero. In general the average difference of pixel values is very vulnerable to slight misalignments in the matching patches. In our patch-based framework, misalignment is the rule rather than the exception. To make our metric robust to (slight) misalignment, we first blur both the test and candidate patches by a Gaussian G with σ = r/9. The score is then