Geodesic pixel neighborhoods for multi-class image segmentation

Multi-class image segmentation is a complex problem that poses several challenges: developing better classifiers, designing more discriminative features, finding efficient optimization techniques and modeling the relations between image pixels in different image regions. In this paper we focus on the last one. A common way to address the problem of structured prediction is to model it as a Conditional Random Field (CRF), but in this paper we take a different approach by using classification and integrating local and global semantic structure constraints directly in the features. Our contribution is threefold. Firstly, we introduce a classification framework based on the concept of pixel neighborhoods, which captures structure constrains with a new histogram based neighborhood feature. Secondly, we propose a novel way to use the geodesic distance to compute the local pixel neighborhood. Thirdly, we introduce a new global rays based neighborhood, again using the geodesic distance, that can also capture global context.