False Contours

The results in Figure 1 show that of the two partitioning (before and after the contour relaxation) the latter is generally more accurate. For instance, the boundary pixel locations in the latter corresponds closely to the edges in the original image (note in particular the regions corresponding to the legs of the tripod). However, an exception to this is the sky region. It is apparent that after the contour relaxation the sky region is broken into many smaller regions. Although this partitioning is mathematically consistent with the underlying model, visually it is not very pleasing. We perceive the sky region (which has a smoothly varying intensity) as one region. In contrast, the segmentation algorithm describes such a region as consisting of many smaller piecewise constant regions. As a result, the boundaries between the smaller regions are perceived as false contours. A method to alleviate this problem is described in section 4.