Implicit Neural Representation For Low-Overhead Graph-Based Holographic-Type Communications

Takuya Fujihashi, Sorachi Kato, Toshiaki Koike-Akino

Abstract: Point cloud delivery over wireless and mobile channels will be a key technology for untethered users to realize extended reality via wireless and mobile terminals. A key challenge of point cloud delivery is efficiently delivering the point cloud over unstable and band-limited channels. Graph Fourier Transform (GFT) is a potential solution to compress such non-uniformly and non-orderly distributed signals in a 3D space, whereas GFT-based solutions require large communication overhead to share the graph information with the receiver. This paper proposes a novel point cloud delivery scheme that introduces implicit neural representation (INR) to reduce the overhead. Specifically, the INR of the proposed scheme trains a mapping between the indices and the corresponding weight in the adjacency matrix and the proposed scheme sends the parameter set of the INR as the metadata. Evaluations demonstrate that the proposed scheme can improve the point cloud quality under the same amount of communication overhead because the proposed INR can predict most of the elements in the adjacency matrix using a small parameter set.