Figure-ground Segmentation using Metrics Adaptation in Level Set Methods A. Denecke 1,2 , I. Ayllon Clemente 1,2 , H. Wersing 2 , J. Eggert 2 , J. J. Steil 1 1- CoR-Lab - Bielefeld University P.O.-Box 10 01 31, D-33501 Bielefeld - Germany {adenecke,iayllon}@cor-lab.uni-bielefeld.de 2- Honda Research Institute Europe Carl-Legien-Str. 30, 63073 Offenbach/Main - Germany Abstract. We present an approach for hypothesis-based image segmen- tation founding on the integration of level set methods and discriminative feature clustering techniques. Building up on previous work, we investigate Localized Generalized Matrix Learning Vector Quantization (LGMLVQ) to train a classifier for fore- and background of an image. Here we ex- tend this concept towards level set segmentation algorithms, where region descriptors are used to adapt the object contour according to the image fea- tures. Finally we demonstrate that the fusion of both methods is capable to outperform their individual applications and improve the performance compared to other state of the art segmentation methods. 1 Introduction In computer vision, figure-ground segmentation (FGS) handles the special case of dividing an image into two regions, containing the object of interest and the background. Hypothesis-driven approaches for FGS rely on an initial hypothesis that provide an a priori assumption (e.g. from user interaction [1] or depth es- timation [2]) about a pixelwise relation to object or background. Unfortunately they typically include incomplete or partially wrong cues which can be caused by the user or algorithmic problems. Hypothesis-based FGS consists of two steps: the modeling of the feature-statistics of the hypothetical fore- and background and the consecutive integration of those statistics in energy minimization tech- niques like Markov random field formulations [1] or level set methods [3]. These algorithms allow for further concepts like interactions of neighboring pixels or contour constraints to derive compact regions. For example, Rother et al.[1] uses Gaussian mixture models together with the min-cut algorithm to optimize the partition of an image. Similarly in [4], histograms are used as region descrip- tors and are integrated into a level set energy functional including a smoothness term to derive compact foreground segmentations. The statistical or descriptive modeling of fore- and background does not respect the discriminability of the used features (e.g. in the case of same colors in fore- and background). In [2], the statistics are modeled with prototypical feature representatives, where an extended learning vector quantization approach [5, 6] is used to train a classifier for fore- and background. There an integrated feature weighting, to discriminate between both regions, is employed.