[Paper] Adaptive Sampling and Layered Compression of Point Cloud Based on Angularity for Human Object Shape Recognition

抄録

Object shapes in physical space are scanned and stored as point clouds for viewing in VR. Since point clouds are represented by huge numbers of points, the data volume becomes enormous, requiring study of appropriate sampling methods and compression coding techniques. When rendering point clouds, if the point cloud lacks attribute information, there is a problem where the backside or internal shapes of the point cloud cannot be seen. We propose the adaptive sampling method to solve this problem. We demonstrate that our proposed adaptive sampling is effective for human object recognition in general scenes containing multiple objects. To efficiently compress the adaptively sampled point cloud, we propose the layered representation extending deep learning-based compression methods. Experiments show that our proposed method can restore the adaptively sampled point cloud with high quality.