To evaluate the performance of
the proposed algorithm in segmenting range images of curved objects, a range image database of a number of objects possessing both planar and curved surfaces from different materials is created. The actual data and their segmented results are shown in the following figures.
Figure 1 - Samples of intensity (left), plotted range data (middle) and segmentation result of the proposed algorithm (right). SSC selected Model 8 (Plane model) for all of the visible sides of the icosahedron.
Figure 2 - Samples of intensity (left), plotted range data (middle) and segmentation result of the proposed algorithm (right). SSC selected model 5 for the cylinder perpendicular to the xy plane (surface 4) and model 3 for the cylinder parallel to the xy plane (surface 5). For surface 6, which is a simple plan, model 8 is chosen by SSC.
Figure 3 - Samples of intensity (left), plotted range data (middle) and segmentation result of the proposed algorithm (right). SSC selected model 5 for the cylinder perpendicular to the xy plane (surface 2) and model 3 for the cylinder parallel to the xy plane (surface 3).
Figure 4 - Samples of intensity (top-left), plotted range data (top-right) and segmentation result of the proposed algorithm (bottom-left) and of UB algorithm (bottom-right). This result illustrates that the proposed algorithm can detect the similarity between two similar cylinders correctly. SSC selected model 3 for the cylinders parallel to the xy plane. The underlying surface models of all planar surfaces are chosen to be model 8 (plane).
Figure 5- Samples of intensity (top-left), plotted range data (top-right) and segmentation result of the proposed algorithm (bottom-left) and of UB algorithm (bottom-right). Although the range data contains substantial noise and invalid data, as shown in the plotted range image, the proposed algorithm has accurately segmented the scene. For surface 13, which is a cylinder perpendicular to the xy plane the model selected is model 5. For all of the planar surfaces, SSC has correctly selected model 8 as the underlying model.
Figure 6 - Samples of intensity (top-left), plotted range data (top-right) and segmentation result of the proposed algorithm (bottom-left) and of UB algorithm (bottom-right). The SSC has selects model 5 for the perpendicular cylinders to the xy plane (surface 16 and surface 15) model 8 for all the planar sections. Surface 17 has two separated planar parts and they are correctly joined together by the proposed technique.
Figure 7- A sample from the K2T range image database: intensity image (left)and segmentation results using the proposed algorithm (right). This image contains cylindrical, general quadratic and planar surfaces.
Figure 8- A sample from the K2T range image database: intensity image (left) and segmentation results using the proposed algorithm (right). The shadow of the doorknob is segmented separately.
Top
By Niloofar Gheissari and Alireza Bab-Hadiashar
May 2004