MDL based Structural Interpretation of Images under Partial Occlusion

In this paper an information theoretic approach is provided for resolving border ambiguity under partial occlusion. The proposed framework allows structural interpretation of images prior to the application of domain specific knowledge. The central idea behind MDL based figure-ground grouping is merging those place-token candidates whose composed description length (whole) is better than sum of their individual description lengths (parts). The computational theory is illustrated by application to blocks-world images.