Sparse Point Cloud Upsampling Based on Neural Implicit Functions

In this paper, we propose a novel point cloud representation method based on neural implicit functions - spatial fields. This method utilizes neural implicit functions to transform three-dimensional coordinate points into local spatial fields, converting the original “discrete-discrete” point cloud representation into a “discrete-continuous” geometric representation, thereby to obtain continuous point cloud representations and richer geometric detail expression. In this method, each three-dimensional coordinate point in the original sparse point set is transformed into a local spatial field embedding multi-layer neighborhood information by means of implicit functions. Eventually, multiple such local spatial fields are aggregated into a continuous high-resolution spatial field to approximate the object surface as closely as possible. At last, arbitrary-scale sampling can be conducted in the high-resolution spatial field for point cloud densification needs at arbitrary resolutions in downstream applications such as 3D medical image reconstruction and autonomous driving. This paper provides an example to illustrate how to utilize the results of the proposed solution for 3D model reconstruction.