Segmentation of 3D scenes using geometric features based on saliency prior for robotic applications

In recent years, there has been increasing interest in how to efficiently and quickly segment objects in robotic tasks. This work offers a supervoxel-based solution to object segmentation in cluttered scenes. Instead of using a uniformly distributed spatial seed resolution, our process uses a salient prior based on scene objects to drive supervoxel generation. Based on a saliency prior, the proposed method first over-segments the 3D scene into multiscale supervoxels with graph structure. The supervoxel maps that have been constructed are then used to distinguish between planar and non-planar structural features. Finally, in an energy optimization framework, we leverage the planar information gleaned from the scene to optimize the original segmentation results. We conducted experiments on OCID and TUW data to verify the effectiveness of our segmentation method.