In this paper, we first generalize a recent statistical color image segmentation algorithm [5] to arbitrary graphs, and report its performance for 2D images, 3D meshes and volume data. We then describe a fast pre-segmentation to the main graph procedure that allows us to further speed-up the segmentation by a factor of ×2 to ×4, without decreasing significantly the quality of segmentations. As an application, we built a realtime video segmentation that is robust enough to be used in camera-driven user interfaces and robotics applications.