HASAN, HOGG: SEGMENTATION USING DEFORMABLE SPATIAL PRIORS 1 Segmentation using Deformable Spatial Priors with Application to Clothing Basela S. Hasan scbsh@leeds.ac.uk David. C. Hogg d.c.hogg@leeds.ac.uk School of Computing University of Leeds Leeds, UK Abstract We present a method for segmenting the parts of multiple instances of a known object category exhibiting large variations in projected shape and colour. The method builds on an existing MRF formulation incorporating a prior shape model and colour distributions for the constituent parts. We propose a novel shape model consisting of a deformable spatial prior probability for the part-label at each pixel. We also make a simple extension to the MRF formulation to deal simultaneously with multiple objects within a global optimisation. Finally, we evaluate the method for the task of segmenting individual items of clothing in images depicting groups of people, and demonstrate improved performance against the state of the art for this task. 1 Introduction Image segmentation is a longstanding problem in computer vision with many potential ap- plications. The formulation of this problem as maximum a posteriori probability (MAP) inference over a Markov Random Field (MRF) is both elegant and effective. Typically, the MRF is configured to favour contiguous regions with the same labelling, and consistency be- tween the label at each pixel and prior intensity distributions for foreground and background regions. Boykov and Jolly show how to solve this labelling problem efficiently by refor- mulating as finding a minimum graph-cut [4]. Finding the min-cut for a given graph can be found by solving an equivalent max-flow problem [2]. The min-cut/max-flow technique provides a globally optimal solution. In an extension to this method, Rother et al. [19] treat the colour distribution for the foreground as a latent property that is optimised along with the labelling in the proposed ”GrabCut“ method. The background distribution is estimated from a user-defined window surrounding the target foreground object. The problem remains that of finding the MAP solution, but now ranging over the space of possible labellings and foreground colour dis- tributions. The method is iterative and finds a local minimum: first initialise the colour distributions from predetermined regions inside the user-defined window; find the optimal segmentation for these initial distributions; then re-estimate the foreground colour distribu- tion from the labelled pixels. This procedure is repeated until convergence. Vicente et al. [23] propose a non-iterative optimization of segmentation and appearance that is shown to outperfom the iterative approach adopted in [19]. c  2010. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms. BMVC 2010 doi:10.5244/C.24.83