Many three-dimensional(3D) object recognition schemes have been proposed. In most cases, it was assumed that 3D database was available. But there are cases where it is not available. In such situations, we must make the database by two-dimensional images of the object taken from all possible views.
Three problems may occur in this case. First, it takes very much time to search all the database. Second, the database may become too large. Finally, it is very difficult to estimate pose if conventional methods are used because they use 3D database for pose estimation.
So, in this paper, we propose another scheme that places pose estimation before the recognition stage. In this scheme, the pose estimation stage selects some candidates out of the whole database by some simple features. The recognition stage makes the final decision from these candidates. From this configuration, we can reduce the overall recognition time.
In addition, an efficient construction of the database is considered. We first determine the number of images that will be stored in the database. It is determined by degree of aliasing caused by small number of samples. And we find angles such that neighboring images have the same similarity. This makes the recognition more accurate at the same number of images.
Simulation results show that the proposed scheme recognizes objects exactly and the overall recognition time is reduced.