Improved Deep Point Cloud Geometry Compression

Abstract

Point clouds have been recognized as a crucial data structure for 3D content and are essential in a number of applications such as virtual and mixed reality, autonomous driving, cultural heritage, etc. In this paper, we propose a set of contributions to improve deep point cloud compression, i.e.: using a scale hyperprior model for entropy coding; employing deeper transforms; a different balancing weight in the focal loss; optimal thresholding for decoding; and sequential model training. In addition, we present an extensive ablation study on the impact of each of these factors, in order to provide a better understanding about why they improve RD performance. An optimal combination of the proposed improvements achieves BD-PSNR gains over G-PCC trisoup and octree of 5.50 (6.48) dB and 6.84 (5.95) dB, respectively, when using the point-to-point (point-to-plane) metric. Code is available at https://github.com/mauriceqch/pcc_geo_cnn_v2.

On Block Prediction For Learning-Based Point Cloud Compression

“…Two early studies are presented in [11,12], using shallow architectures composed of convolution and de-convolution layers for analysis and synthesis. The impact of several parameters added to the initial version of the former network is evaluated in [13]. In [14], a rate-distortion performance analysis is conducted on the latent space of [12], which is enriched with a hyper-prior and the possibility of explicit quantization in [15].…”

A Deep Point Cloud Geometry Coding Toolbox

“…This open source package offers a toolbox to compress point cloud geometry using deep neural networks. It implements all the contributions and experiments in our previous work [4]. The point cloud is partitioned into blocks and each block is compressed with a convolutional neural network, optimized with respect to a focal loss.…”

“…At the decoder side, several parameters can be modified to adapt to the local point cloud density (balancing weight in the focal loss, optimal thresholding). Details on the choice of these parameters are reported in [4]. We complete the toolbox with a number of utility functions for evaluating and comparing the performance with other methods; we also provide pre-trained models and the datasets used to obtain them.…”

“…This open source package is intended for researchers and engineers working on PCC. It has two main purposes: i) facilitating the reproduction and comparison with the family of codecs in [4]; ii) providing the tools and a starting point to foster new research on deep PCC. In particular, the adopted licensing scheme allows the use of this software as a base for implementation in an industrial context.…”

Deep Learning-Based Quality Assessment Of 3d Point Clouds Without Reference

“…been an increasing interest in the multimedia research community towards the definition of accurate quality metrics for point clouds. These efforts have been also partially motivated by the concurrent standardization activities for point cloud coding in MPEG [1], as well as due to the need of relevant loss functions in learning-based PC compression methods [2,3]. Different point cloud objective metrics have been proposed in the literature.…”