Department of Image Processing

Objectives

Recognition of partially occluded objects is an important task in a general 3D scene understanding. We focused our research to recognition of binary 2D objects with complicated curved boundary. Such a boundary cannot be approximated by a polygon without loss of accuracy and the object cannot be identified by its surface. We model the objects deformation by an unknown affine deformation. It was proved in our experiments that the affine transformation is a good approximation for analyze of 2D images on a 3D scene.

Method

A new method has been developed. It consists of three steps:

1. Control point detection. The control points are affine invariant point with a zero curvature – inflection points or central parts of straight lines.
   

   

2. Description of object by parts. The object contour is to parts, separated by the control points. The parts are by a modified radial vector, which is invariant to affine transformation and is robust to non-convex objects.
3. Recognition is realized by a kind of string matching. We use the invariant descriptors as well as recognition based on the object transformation.

Results

Our experiment proved a good discrimination power of the developed method. The tests on 24 objects database, test on noisy data and test on real images have been done. We have found that the principal limitation remains the stability of the inflection points.

Other application

Our curvature estimation (used in step one) has found its use in psychological research. We have generated images with minimum of contour points so that the image could be recognized by a human.

Details:

Publications:

• 2008 – Horáček Ondřej, Kamenický Jan, Flusser Jan: Recognition of partially occluded and deformed binary objects , Pattern Recognition Letters vol.29, 3 (2008), p. 360-369
• 2005 – Horáček Ondřej, Kamenický Jan, Flusser Jan: Recognition of partially occluded and deformed binary objects , Proceedings of the 11th International Conference on Computer Analysis of Images and Patterns. CAIP 2005 vol.3691, - (2005), p. 415-422, Computer Analysis of Images and Patterns /11./, (Versailles, FR, 05.09.2005-08.09.2005)
• 2004 – Horáček Ondřej: Diploma Thesis - Rozpoznávání částečně zakrytých objektů s lokálním využitím afinních invariantů, 2004, ČVUT FJFI, Prague, ČR
• 2002 – Horáček Ondřej: Research project - Rozpoznávání částečně zakrytých objektů, 2002, ČVUT FJFI, Prague, ČR