A Supervoxel Segmentation Method With Adaptive Centroid Initialization for Point Clouds

Supervoxels find applications as a pre-processing step in many image processing problems due to their ability present regional representation of points by correlating them into set clusters. Besides reducing the overall computational time for subsequent algorithms, desirable properties supervoxels are adherence object boundaries and compactness. Existing supervoxel segmentation methods define size based on user inputted resolution value. A fixed results poor performance point clouds with non-uniform density. Whereas, other methods, quest better boundary adherence, produce irregular shapes elongated boundaries. In this article, we propose new method, k-means algorithm, dynamic cluster seed initialization ensure uniform distribution seeds variable densities. We also strategy, histogram binning surface normals, adherence. Our algorithm is parameter-free gives equal importance color, spatial location orientation resulting compact tight test efficacy our publicly available cloud dataset consisting 1449 pairs indoor RGB-D images, i.e., color (RGB) images coupled depth information (D) mapped per pixel. Results compared against three state-of-the-art algorithms four quality metrics. show that method provides significant improvement over undersegmentation error compactness metrics and, performs equally well recall contour density