Optimization and Tracking of Polygon Vertices for Shape Coding Janez Zaletelj, Jurij F. Tasic University of Ljubljana, Faculty of Electrical Engineering, Trzaska 25, SI-1000 Ljubljana, Slovenia {janez.zaletelj, jurij.tasic}@fe.uni-lj.si http://ldos.fe.uni-lj.si Abstract. The efficiency of shape coding is an important problem in low-bitrate object-based video compression. Lossy contour coding meth- ods typically include contour approximation by polygons or splines, spa- tial and/or temporal prediction of vertices, and entropy coding the of prediction error. In conventional contour coding schemes, however, the coding gain in the interframe mode is typically small. This indicates that the temporal redundancy is not successfully removed. The paper addresses the issue of temporal shape decorrelation by proposing the Kalman filtering-based approach to vertex tracking and prediction. A temporal vertex association procedure is proposed effectively minimiz- ing bit rate in each frame. The prediction error is coded using adaptive arithmetic encoding. Vertex optimization is employed to reduce the shape reconstruction error. 1 Introduction In the context of object-based video coding, shape information is crucial for content-based access and manipulation of video streams. Because it represents a substantial part of the total bit rate, efficient coding methods are needed. MPEG-4 uses Context Arithmetic Encoding [9], which is a bitmap-based cod- ing method. However, lossy contour-based coding methods can achieve a higher coding efficiency by encoding control points / vertices of spline or polygon ap- proximation. Most contour-based shape coding methods [1,2,3,5,10,11] concen- trate on finding a set of polygon vertices / spline control points to minimize the bit rate while satisfying the maximum allowable distortion criterion. Intra-frame relative addressing of vertices is employed and predefined variable length codes are used for entropy coding, which limits the rate-distortion efficiency. Temporal decorrelation of shape is generally not addressed adequately, with the exception of [7] which uses Lagrangian optimization of control point positions to obtain a rate-distortion optimal solution. The computational complexity of the optimal methods [5], [7] is quadratic in the number of admissible control points. Their optimality can only be claimed within the limitations imposed by the chosen code structure, motion compensa- tion scheme, approximation scheme, width of the admissible control-point band,