Point cloud upsampling using deep self-sampling with point saliency

Point cloud upsampling is a process of increasing the point density to represent an object or environment effectively. Recent studies have focused on deep learning-based approaches that learn mapping from a sparse to a dense region of the point cloud. Self-supervised learning-based upsampling techniques have gained attention due to their capability to learn predefined characteristics without previous training on a large dataset. This study proposes deep self-sampling with point saliency. The approach involves the use of a self-sampling network with two predefined consolidation strategies, namely density and curvature, along with a saliency feature, to restore the underlying characteristics of an object effectively. Additionally, multistep upsampling is applied to determine the best order of different consolidation strategies for optimal results. Experimental results show that multistep self-sampling using point saliency outperforms the existing approach because it can effectively restore the underlying shapes of the object qualitatively and quantitatively.