Part{based Decomposition from Edge Images: Human Responses vs. Mdl Segmentation

This paper is concerned with an aspect of the structural properties of images of simple objects, namely their decomposition into generic parts using edge information only. In particular, it reports results of a psychological experiment conducted using experimentally naive subjects to assess their concept of part when presented with poor edge images. It compares the results with those of a recent computational method based on generic deformable part models and Minimum Description Length (MDL) segmentation of the scene in terms of these parts. The outcome shows a certain degree of convergence but disparity amongst subjects resemble di culties of computational approaches.