Unseen Object Instance Segmentation for Robotic Environments

In order to function in unstructured environments, robots need the ability recognize unseen objects. We take a step this direction by tackling problem of segmenting object instances tabletop environments. However, type large-scale real-world dataset required for task typically does not exist most robotic settings, which motivates use synthetic data. Our proposed method, instance segmentation (UOIS)-Net, separately leverages RGB and depth segmentation. UOIS-Net is composed two stages: first, it operates only on produce center votes 2D or 3D assembles them into rough initial masks. Second, these masks are refined using RGB. Surprisingly, our framework able learn from RGB-D data where nonphotorealistic. To train we introduce random objects tabletops. show that method can sharp accurate masks, outperforming state-of-the-art methods also segment robot grasping.