Usually, a pixel of a digital image represents the reflected color or grey tone of a given projected point of the real space onto a charge coupled device (CCD), or similar sensor cell. However, range images are those in which every pixel of the image contains information about the distance from that point to the observing sensor. Typically, these sensors are long-range laser scanners. Before the use of such sensors, range images adopted a quite simple approximation, based on basic polyhedral objects. New challenges have now emerged; now, range objects are complex, and have very different sizes.
The authors of this paper propose an algorithm for segmenting three-dimensional (3D) scenes and reconstructing object surfaces from range images. This method explores the solution space with two types of moves: reversible jumps and stochastic (Bayesian) diffusions. The authors note six types of moves when exploring the solution space; there are six types of different jump and diffusion strategies.
The algorithm proposed by the authors is tested using three different range data sets, and the results are very interesting. Many different techniques are used and integrated. Researchers studying the segmentation and clustering of range images should take this contribution into account. However, this paper is not suitable for a beginner, because, although it is quite well written and clear, it is hard to follow. I have one minor quibble: the processing time is noted to be about one hour; further experiments involving time consumption, and more information about that, should have been included.