Foldingnet-Based Geometry Compression of Point Cloud with Multi Descriptions

Abstract

Traditional point cloud compression (PCC) methods are not effective at extremely low bit rate scenarios because of the uniform quantization. Although learning-based PCC approaches can achieve superior compression performance, they need to train multiple models for different bit rate, which greatly increases the training complexity and memory storage. To tackle these challenges, a novel FoldingNet-based Point Cloud Geometry Compression (FN-PCGC) framework is proposed in this paper. Firstly, the point cloud is divided into several descriptions by a Multiple-Description Generation (MDG) module. Then a point-based Auto-Encoder with the Multi-scale Feature Extraction (MFE) is introduced to compress all the descriptions. Experimental results show that the proposed method outperforms the MPEG G-PCC and Draco with about 30% ~ 80% gain on average.