Alignment of 2D objects for shape interpretation

Abstract

Humans usually describe objects along a certain direction, called intuitive direction, in other words place them in a way that they are commonly seen in their surroundings. In computer vision, the intuitive alignment may be very useful for object interpretation and semantic classification. For example, it may facilitate the extraction of characteristic points such as the base and apex of plant leaves, the eyes and tail of fishes and so on, which enrich greatly the representation of their shapes. While it still an open challenge to automatically align objects along their intuitive orientation, this paper examines opportunities to determine it for objects that have a somewhat symmetric shapes. Inspired by an idea related to 3D alignment, our approach is based on two types of symmetry: reflectional symmetry and local translational symmetry. Experimental results carried on the MPEG-7 dataset show that our method detects an alignment that corresponds to the users' intuition for most objects that have an approximately or partially symmetric shapes.