Complement decoded point cloud with coordinate adjustment for video-based point cloud compression

Dynamic point cloud (DPC) represents a realistic 3D scene in motion and has a wide range of applications. Compressing point clouds has become crucial for storing and transmitting such data. Video-based point cloud compression (V-PCC) developed by the Moving Picture Expert Group can achieve remarkable performance in DPC compression. However, it also introduces issues of point reduction and coordinate distortion in the decoded DPC. In this paper, we present a 3D-based framework as a post-processing tool for the V-PCC decoder, which complements decoded DPC and performs coordinate adjustment. In particular, we propose a neighbor-based interpolation method to recover the missing points based on the coordinates in decoded DPC. Then, to minimize the coordinate distortion in interpolation, we design a sparse fully convolutional networks, 3D Minkowski Unet, to perform coordinate adjustment. Considering the variation of data size for DPC, we propose a cube-based patch generation method to enable the scalability of the proposed framework. The experiment results demonstrate that the proposed framework obtains significant performance in complementing reduced coordinates in both objective and subjective evaluation.