Sparse Fully Convolutional Network for Video-based Point Cloud Compression Color Enhancement

Dynamic point cloud enables objects or scenes to have a realistic 3D representation in motion. Storage and transmission of dynamic point cloud efficiently is an essential precondition for its application. Video-based point cloud compression (V-PCC) developed by the MPEG standardization group can achieve remarkable performance in compressing dynamic point clouds. However, it also introduces compression noise in decoded dynamic point clouds, which can significantly affect subsequent applications. In this paper, we propose a quality enhancement architecture that focuses on improving color attributes on V-PCC compressed point cloud. The architecture designs a sparse fully convolution networks using Minkowski Engine to maintain the sparsity nature of point cloud data and speed up the learning process with less memory usage. Additionally, we applied a feature extraction unit that takes into account the information across channels. Considering the influence of coordinates compression noise on our architecture and the limitation of GPU memory capacity, coordinates optimization and patch generation methods are applied to input data as a pre-processing step. To the best of our knowledge, this is the first implementation of the Minkowski Engine for enhancing color attributes of compressed point clouds in the V-PCC field. The experiment results demonstrate that the proposed architecture can improve the quality of color attributes in the reconstructed point cloud with different quantization parameters.