An experimental comparison of Extended Gaussian Image and Shape Distributions in 3D shape retrieval

摘要

The effective retrieval of three-dimensional shapes is a very crucial problem nowadays. It has to be not only efficient but also carried out in reasonable time. The last demand is especially difficult as 3D objects are usually built using lots of data (vertices, patches, etc.). That was the reason for minor interest dedicated few decades ago by scientists to them. At present, this problem became less important, thanks to the advances in computer hardware development. Now, one can find many new applications of 3D models, e.g. in CAD systems, entertainment, virtual reality, biometrics and image retrieval. In order to work with those objects three-dimensional shape descriptors are used. Those algorithms are created to represent objects independently of various problems concerning them, e.g. affine transformations, noise, occlusions. The result of experimental examination of two 3D shape descriptors is provided in the paper. The research was performed using the models from the "Princeton Shape Benchmark" database. This database is very popular in the task of experimental evaluation of 3D shape descriptors. In the paper two methods of that type are explored - Extended Gaussian Image and Shape Distributions-in the problem of 3D shape retrieval.