PROGRESSIVE POLYGON ENCODING OF SEGMENTATION MAPS Marc Servais and Theo Vlachos CVSSP, University of Surrey Guildford, GU2 7XH, United Kingdom m.servais@surrey.ac.uk Thomas Davies BBC Research and Development Kingswood Warren, Tadworth, Surrey, KT20 6NP, United Kingdom ABSTRACT A new method for the lossy compression of segmentation maps is presented. In the first stage of the proposed al- gorithm, region boundary segments are coded, providing a coarse polygonal approximation of regions, as well as a complete representation of their adjacency. The second stage consists of the progressive refinement of these seg- ments, until either the desired degree of accuracy is reached, or the target bit-rate is achieved. Rate-distortion curves de- monstrate good lossy performance in the range of 0.3 to 0.6 bits per contour point. 1. INTRODUCTION Segmentation maps are employed in a variety of applica- tions, including object-based video processing and coding, where they are used to define a region-based description of an image. Every region has a unique label and consists of a connected neighbourhood of pixels. It is often desirable to compress segmentation maps as they contain a high de- gree of redundancy. An important consideration is how to achieve the right balance between the degree of compres- sion on the one hand and the fidelity of the decoded map on the other. Shape compression (for individual foreground objects) has been researched extensively in recent years and two main approaches have emerged: direct bitmap coding of the object mask, and coding of the shape boundary. The latter method has been shown to allow a greater degree of com- pression [1]. Region boundary coding is typically achieved through the use of (lossless) chain-coding or (lossy) poly- gon/spline approximation [2]. In the case of lossy region boundary coding (with poly- gons or splines), the goal is to minimise both the coding cost and the distortion caused by approximating the boundary [2, 3, 4]. Some approaches impose additional criteria, such as minimising the number of polygon vertices for a given permissible error [5]. Progressive polygon encoding [6] has This work has been funded by BBC Research and Development and the Centre for Vision Speech and Signal Processing, University of Surrey. been proposed as a way of enabling progressive transmis- sion of a region’s shape. This can proceed until either the desired bit-rate is reached, or the error is sufficiently small. The methods outlined above are for individual regions (e.g. video objects in MPEG-4). However, a segmentation map can be viewed as a collection of regions. This suggests that many of the approaches used in shape coding would be useful in the coding of segmentation maps. Furthermore, the coding cost can be reduced because common boundaries between neighbouring regions only need to be coded once. Some research into the coding of segmentation maps has been reported within the context of region-based video cod- ing [1, 7, 8]. However, in these cases the aim was to min- imise the global coding cost, thus not necessarily achieving an optimally coded segmentation map. This paper describes a novel algorithm for the progres- sive encoding of segmentation maps. First, common bound- ary segments between neighbouring regions are identified. Using these segments an initial polygon structure is cre- ated, which only coarsely approximates the actual map, but which fully specifies the connectivity between neighbouring regions (Section 2). Following this, the existing segments are progressively refined until either the desired bit-rate is reached, or an acceptable error is achieved (Section 3). Finally, experimental results are presented for two seg- mentation maps encoded at a variety of bit-rates (Section 4). The performance of the proposed method and the ad- vantages it offers to region-based video coding schemes are discussed. 2. ENCODING INITIAL POLYGON STRUCTURE Given a segmentation map comprising a number of regions, the goal is to approximate each region with a polygon. How- ever, a simple polygon approximation of each region in turn is likely to result in a new segmentation map in which the the polygon-shaped regions are not correctly aligned. This is because the polygons corresponding to neighbouring re- gions could either overlap or leave gaps along their common boundary. This can be avoided by the method outlined be- low and illustrated in Figure 1: 0-7803-8554-3/04/$20.00 ©2004 IEEE. 1121