A shape-based camera angle-invariant retrieval scheme for 3D objects

In many applications including security, surveillance and place recognition, camera angle-invariant object retrieval is required. Generally, for shape-based angle-invariant image retrieval, we need an object image at every angle where the shape of the image changes. Moreover, if we want to guarantee image rotation, we need to consider all possible shapes resulting from image rotation. These two rotations will generate a huge image database. If we simply apply existing indexing schemes to this database, serious performance problems will occur due to database size and redundancy. In this paper, we propose a new indexing and matching scheme for shape-based camera angleinvariant image retrieval, focusing on two types of camera movements: horizontal rotation and fixed rotation. To reduce the space requirement for indexing a huge database, we used two levels of indexing structures: static indexing for horizontally rotated images for horizontal angle invariance, and dynamic indexing for fixedly rotated images for rotation invariance. Based on the proposed indexing scheme, we implemented a prototype image retrieval system. We showed by experiment that our proposed scheme can achieve much better performance than competing methods.