Planar Shape Databases with Affine Invariant Search

Image databases are often used to archive and retrieve images containing man-made 3D objects usually taken from arbitrary viewpoints. These objects generally incorporate planar surfaces containing different kinds of highly curved patterns. It is often the case that the form of such patterns characterizes well the corresponding object. Besides classical retrieval by colour or texture, the database should therefore inevitably support queries based on geometrical shape. In this paper we show how images in a database can be provided with additional information allowing viewpoint invariant shape query. Viewpoint invariance allows to compensate for affine or projective distortions, and is achieved by the usage of invariant reference frames. Those frames serve as a canonical coordinate system, common for all shapes in the database; in this system their similarity can be estimated and neighbourhood relationships can be established for subsequent search. The contributions of this work are twofold: first, the construction of such reference frames is done by combining different types of local invariant properties of the curve. Second, in the context of search operations, an important speed-up is introduced by means of momentbased indexing. Both strategies are tested on a database of trademarks. The database contains images of supermarket box packages with trademarks placed on their flat sides. Section 2 briefly describes the method for computing viewpoint-independent shape similarity. Section 3 presents the search mechanism, including indexing speed-up routine, and an evaluation of the results.