Image noise is an important issue in vision computing. We created several synthetic images with all primitive surfaces (plane, cylinder, ellipsoid and hyperboloid). The test data were formed by adding uncorrelated Gaussian distributed noise with zero mean and variable standard deviation (for the different noise levels) to the images. Figure 3 shows the minimum percentage of mislabelled points for the best possible threshold for a certain noise level.
We did one test with an image that contains equal large plane, cylinder, ellipsoid and hyperboloid surfaces. We performed the experiment with one test image that contains the four shapes and another test images which contains the four shape regions but the three curved surfaces are each replaced by four different cylinder, ellipsoid and hyperboloid surfaces. The SC segmentation performs slightly better in the first test image (see Figure 3 left). The methods perform differently in the second test image (see Figure 3 right), where the HK algorithm is affected by the noise at much lower levels than in the single surface images. The performance of the SC method is significantly better over all noise levels.
Figure 3: Mislabelled points versus noise level