Exploiting Protrusion Cues for Fast and Effective Shape Modeling via Ellipses

Modeling objects with a set of geometric primitives is a key problem in computer vision and pattern recognition with numerous applications including object detection and retrieval, tracking, motion and action analysis. In this paper, we attempt to represent 2D object shapes with a number of ellipses with semantic meaning (e.g. one ellipse may correspond to an arm, while two other ellipses may represent a bent leg), while maintaining a high coverage of the shapes. We propose a novel ellipse fitting method based on psychology and cognitive science studies on shape decomposition and show that our shape coverage compares well with the state of the art methods, while significantly outperforming them in run-time by as much as 508 times in our evaluation of the methods on over 4000 2D shapes. We also demonstrate the value of our method to higher-level processing via an example application in creating 3D ellipsoid models of PASCAL horses.