Scale adaptive supervoxel segmentation of RGB-D image

Superpixels are perceptually meaningful atomic regions that can effectively capture image features. We propose a novel scale adaptive supervoxel segmentation algorithm for RGB-D images, i.e., small supervoxels in content-dense regions (e.g., with high intensity or color variation) and large supervoxels in content-sparse regions. Among various methods for computing uniform superpixels, simple linear iterative clustering (SLIC) is popular due to its simplicity and high performance. We extend SLIC to generate small evenly distributed supervoxels in 3D space as PDS-SLIC, reducing the numbers of nodes. Then we use supervoxels as the nodes to construct a graph and setting the adaptive threshold for supervoxel merging. Experiments on NYU Depth Dataset V2 show that our proposed method outperforms state-of-the-art methods.