CBIR over multiple projections of 3D objects

This paper presents a heuristic approach to 3D object recognition by considering multiple 2D projections (appearances) of the objects of interest. Thus, 3D object identification is interpreted as a conventional Content Based Image Retrieval (CBIR) problem. An arbitrary input image of a given object is treated as a search sample within a database (DB) of a large enough set of images, i.e. appearances from a sufficient number of viewpoints for each object. The CBIR method to access the image DB should be both fast enough and sufficiently noise-tolerant. The method we propose is described over two cases of recognition, namely human faces and hand signs of a given sign-language alphabet. Analogically, the method can also be applied to recognition of a large number of 3D objects of different types. We are briefly covering the data gathering technique, its structuring into a DB of image samples, and the experimental study for the noise-resistance of the applied CBIR method. The latter is used to acknowledge the applicability of the proposed approach.