Supervoxel-based 3D Segmentation in Complex Scenes

We present an unsupervised 3D point cloud segmentation method for household table-top scenes, which are usually unstructured and highly cluttered. The main idea of the algorithm is based on graph-cut which conducts upward merging steps from roar point cloud to adjacent supervoxles, homogeneous surface patches and to final object hypotheses. Our approach based only on 3D point cloud which contains more accurate spatial information than 2.5D images. The proposed approach together with other two state-of-the-art methods are compared and evaluated on a benchmark database. Advantages and weaknesses of our method are also discussed in the end.