Adaptive Geometry Partition for Point Cloud Compression

Octree (OT) geometry partitioning has been acknowledged as an efficient representation in state-of-the-art point cloud compression (PCC) schemes. In this work, an adaptive geometry partition and coding scheme is proposed to improve the OT based coding framework. First, quad-tree (QT) and binary-tree (BT) partitions are introduced as alternative geometry partition modes for the first time under the context of OT-based point cloud compression. The adaptive geometry partition scheme enables flexible three-dimensional (3D) space representations and higher coding efficiency. However, exhaustive searching for the optimal partition from all possible combinations of OT, QT and BT is impractical because the entire search space could be huge. Therefore, two hyper-parameters are introduced to specify the conditions on which QT and BT partitions will be applied. Once the two parameters are determined, the partition mode can be derived according to the geometry shape of current coding node. To investigate the impact of different partition combinations on the coding gains, we conduct thorough mathematical and experimental analyses. Based on the analyses, an adaptive parameter selection scheme is presented to optimize the coding efficiency adaptively, where multi-resolution features are extracted from the partition pyramid and a decision tree model is trained for the optimal hyper-parameters. The proposed adaptive geometry partition scheme has shown significant coding gains, and it has been adopted in the state-of-the-art MPEG Geometry based PCC (G-PCC) standard. For the sparser point clouds, the bit savings are up to 10.8% and 3.5% for lossy and lossless geometry coding without significant complexity increment.

I. Introduction

Point clouds have recently been recognized as a significant media format, because it reveals the essential geometry structure of three dimensional (3D) scenes and has potentials in a wide variety of applications. These applications can be classified into two categories in terms of the ultimate utility and receptors. The first one is human-oriented applications such as virtual reality (VR), augmented reality (AR), culture heritage etc., which intend to provide immersive, photo-realistic and interactive viewing experiences for human eyes. One recent representative work in this category uses point clouds as the representation of view-dependent light fields with unlimited six-degrees-of-freedom [1], [2]. The second category is machine-oriented applications such as autonomous navigation for vehicles or drones, where the ultimate receptors in this category are machines rather than human eyes. Machine-oriented applications focus on object detection, object segmentation, localization and other computer vision related tasks, in which great progresses have been witnessed recently on manipulating 3D point clouds by deep learning based methods with neural networks [3]–[6].