Describing and recognizing 3-D objects using surface properties

From MaRDI portal





The problem attacked is the following. Suppose that we obtain ``range data from a scene consisting of a number of three-dimensional objects from a specified set. (Range data is a function of two variables which is physically obtained as follows. A rectangular planar segment, referred to as the ``viewing plane, is placed in space so that the union of the set of lines intersecting the viewing plane perpendicular to it intersects the objects in the scene. The value assigned to any point in the viewing plane is the distance from it to the nearest object surface in the scene.) We desire to recognize which objects in the specified set are in the scene. The recognition is based on descriptions of the objects in the specified set consisting of range data for the individual objects for a few (2-6) different viewing planes. A complete system is described to do this. The system is demonstrated on range data of scenes consisting of up to four objects, some occluding the others. The objects in these demonstrations are typically simplified models of real objects, such as a telephone or a car. The system appears to work remarkably well in this artificial environment. The essential steps taken by the system are the following. 1. Identification of images of individual surfaces in the range data and the fitting of second-order polynomial surface patches to them. 2. Representation of the objects in the range data by ``attributed graphs. The nodes of such a graph correspond to individual surface patches, the arcs describe adjacency relations between the patches, and the attributes assigned to them are geometrical properties calculated from the range data. 3. Recognition by matching these attributed graphs to those of the given range data for the individual objects. The author is to be congratulated for having developed, implemented, and tested a complete (and by necessity complicated) system to achieve the stated task. On the other hand, the presentation of the material leaves a lot to be desired. It has the style of a doctoral dissertation whose purpose is to impress the committee with knowledge acquired and work performed rather than to educate its reader. For example, there is no index. Also, approaches alternative to the authors are briefly mentioned (without sufficient explanation to be understandable) and then criticized without evidence supporting the criticisms. This reviewer was not convinced that the author's approach is indeed superior to the alternative in the literature. The whole of Chapter 2 (Survey of previous work) is probably quite useless to those not working in the field. Unfortunately the same criticism can also be leveled to much of the rest. For example, Section 3.1 starts with the sentence: ``It is well known in differential geometry that a surface can be reconstructed up to second order from the knowledge of curvature at each point, except for a constant term (by using the first and second fundamental form, see for example). The terms used in this sentence are not defined (not even ``curvature``). Is the reader, presumably a computer scientist, supposed to first read a book on differential geometry to follow the work under review? With a little effort the author could have provided the necessary definitions and could have quoted, without proofs, the necessary theorems. Another problem from the point of view of educating the reader is the way the rules used by the system are presented. While they are described in detail so that another person could duplicate most of the system based on the book, the reasons for the specific values used by the system are not explained. Here is one of many potential examples: ``If any of the following tests fails, \({\mathbb{M}}\) and S are considered not similar enough and the model view \({\mathbb{M}}\) is discarded: \(N({\mathbb{M}},S)>40 \%\), \(R({\mathbb{M}},S)>30 \%\), \(A({\mathbb{M}},S)>30 \%.\) No reasons are given for the choice of these percentages. In summary, this book reports on some interesting and, by the standards of 3-D object recognition, quite successful work, but the educational value of the book could have been much improved by better presentation.











This page was built for publication: Describing and recognizing 3-D objects using surface properties

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3995274)