Real-time pixel-wise grasp affordance prediction based on multi-scale context information fusion