The performance of both methods on images containing single surfaces is basically the same. One difference is that the HK segmentation is using the zero-threshold to recognise cylindrical surfaces. Within the HK segmentation cylinder points vanish from the image at low thresholds. Therefore, the SC classification is more stable at low thresholds on scenes containing cylinders. For the SC algorithm a slightly higher threshold should be used for the same error rate as the HK algorithm (but this effect is largely unimportant in terms of classification performance). In our noise tests, the SC algorithm can deal better with image noise in images which contain different surfaces, because the HK segmentation cannot focus with an optimal threshold on a single surface. Thus we conclude Koenderink's SC classification scheme has a slight advantage (5-10% lower error rate) when dealing with real scenes containing multiple surfaces and moderate noise.