Hierarchical Prior-based Super Resolution for Point Cloud Geometry Compression

TL;DR

The paper tackles distortions from naive geometry quantization in lossy G-PCC by introducing HPSR-PCGC, an encoder-side hierarchical prior that enables coarse-to-fine decoder-side super resolution for point cloud geometry. By constructing a multi-level prior from a downsampled point cloud pyramid and encoding both the base geometry and priors losslessly, the method achieves substantial rate-distortion gains over G-PCC implementations and SRLUT, while maintaining competitive runtimes. The approach narrows the gap toward V-PCC and PCGCv2 on solid clouds and offers a practical path toward density-adaptive, jointly compressed geometry and attributes in future work. Overall, HPSR-PCGC provides a principled, encoder-centric mechanism to leverage cross-scale geometry features for improved point cloud compression.

Abstract

The Geometry-based Point Cloud Compression (G-PCC) has been developed by the Moving Picture Experts Group to compress point clouds. In its lossy mode, the reconstructed point cloud by G-PCC often suffers from noticeable distortions due to the naïve geometry quantization (i.e., grid downsampling). This paper proposes a hierarchical prior-based super resolution method for point cloud geometry compression. The content-dependent hierarchical prior is constructed at the encoder side, which enables coarse-to-fine super resolution of the point cloud geometry at the decoder side. A more accurate prior generally yields improved reconstruction performance, at the cost of increased bits required to encode this side information. With a proper balance between prior accuracy and bit consumption, the proposed method demonstrates substantial Bjontegaard-delta bitrate savings on the MPEG Cat1A dataset, surpassing the octree-based and trisoup-based G-PCC v14. We provide our implementations for reproducible research at https://github.com/lidq92/mpeg-pcc-tmc13.

Figures (14)

• Figure 1: System diagram of the proposed hierarchical prior-based super resolution for point cloud geometry compression.
• Figure 2: Illustration of a point cloud pyramid produced by the successive downsampling. ${\bm V}, {\bm V}^{(0)}, {\bm V}^{(1)}$, and ${\bm V}^{(2)}$ are shown from left to right.
• Figure 3: Pipeline of the hierarchical prior construction.
• Figure 4: 2D illustration for constructing the interpolation patterns ${\bm \sigma}^{(K)}$ that help map ${\bm V}^{(K)}$ to an approximation of ${\bm V}^{(K-1)}$, where $2^L\times q=3/4$ and the neighborhood consists of the left and right voxels only. Gray/white squares indicate occupied/void voxels in ${\bm V}^{(K)}$, while gray/white circles indicate occupied/void voxels in ${\bm V}^{(K-1)}$. With a factor of $3/4$, points denoted by gray circles are downsampled to the same point denoted by the circumscribed square. "Stats." indicates simple frequency-based statistical analysis.
• Figure 5: 2D illustration for constructing the interpolation patterns ${\bm \sigma}^{(k)}$ where $k=K-1,\cdots,1$ that help map ${\hat{{\bm V}}}^{(k)}$ to an approximation of ${\bm V}^{(k-1)}$. Only left and right neighbors are considered. Gray/white squares indicate occupied/void voxels in ${\hat{{\bm V}}}^{(k)}$, and gray/white circles indicate occupied/void voxels in ${\bm V}^{(k-1)}$. When ${\bm V}^{(k-1)}$ is downsampled with a factor of $1/2$, points denoted by gray circles are merged to the same point denoted by the circumscribed square. We first partition ${\hat{{\bm V}}}^{(k)}$ into several clusters $\{{\hat{{\bm V}}}^{(k)}_{r}\}$ based on neighborhood information, and then obtain the prior ${\bm \sigma}^{(k)}$ based on frequency statistics.

Theorems & Definitions (1)

• proof