GPCGC: A Green Point Cloud Geometry Coding Method

Qingyang Zhou, Shan Liu, C.-c. Jay KuoICIP 2023 · DOI

TL;DR

This paper introduces GPCGC, a novel point cloud geometry compression method that achieves comparable coding performance to existing standards like MPEG's V-PCC, but with significantly lower computational complexity. It achieves this by combining an octree-based hierarchical structure with a low-complexity Green Image Codec (GIC) for depth map encoding, offering a more efficient solution for static point cloud data.

What Problem Does This Paper Address?

Point clouds, essential for applications like virtual reality, autonomous driving, and 3D modeling, generate massive amounts of data. Efficiently storing and transmitting this data, particularly the 3D geometry, is a significant challenge. While existing standards like MPEG's V-PCC and G-PCC offer good compression, they can be computationally intensive. Emerging deep learning-based methods achieve impressive gains but at the cost of even higher complexity and large model sizes. This paper aims to bridge this gap by developing a point cloud geometry compression method that is both computationally efficient and achieves competitive coding performance, specifically for static point clouds.

Methods

GPCGC is built upon two core modules:

  1. Octree Split and Projection:
    • Hierarchical Octree Structure: Unlike V-PCC's patch-based approach, GPCGC first voxelizes the point cloud and then recursively splits it into an octree. This hierarchical decomposition allows for a coarse-to-fine representation.
  2. 3D-to-2D Projection: Projections are performed only at the leaf nodes of the octree. The method selects the projection axis (x, y, or z) that yields the largest projection area to simplify subsequent encoding.
  3. Depth Map Generation: The distances of points to the chosen projection plane form a depth map. To handle projection ambiguities (self-occlusions, multiple points mapping to the same 2D block), GPCGC generates two depth maps (max and min distances) and concatenates them.
  4. Lossy Occupancy Map Coding: Instead of lossless encoding like V-PCC, GPCGC uses a lossy approach for occupancy maps, classifying blocks into 9 modes. This significantly reduces bit rate at the cost of minor distortion, which can be mitigated by depth map inference.
  5. Block Occupancy and Depth Coding:
    • Green Image Codec (GIC): The depth maps, treated as grayscale images, are encoded using the Green Image Codec (GIC). GIC is a multi-resolution codec based on Vector Quantization (VQ) and is designed for significantly lower complexity than traditional video codecs like HEVC-Intra.
  6. Rate-Distortion Optimization (RDO): GPCGC employs RDO to determine the optimal octree split level, balancing coding rate and distortion. This is a key differentiator from many deep learning-based methods that often have fixed rate-distortion trade-offs.

Datasets

  • Training: Mesh models from ShapeNet and SHREC'19 were used. These were converted to point clouds, voxelized, and then split into cubes of various sizes for training.
  • Testing: Single frames from the 8iVFB and Owlii point cloud datasets were used, following MPEG Common Test Conditions.

Key Insights and Ideas

  • Octree Projection as a Sweet Spot: By projecting only at octree leaf nodes, GPCGC avoids the complexity of global atlas generation while still leveraging the hierarchical structure for efficient representation.
  • Lossy Occupancy for Bitrate Savings: The decision to use lossy occupancy map coding is a crucial insight. It demonstrates that sacrificing perfect occupancy information can lead to substantial bitrate reductions without severely impacting overall quality, especially when combined with depth map inference.
  • GIC's Low-Complexity Advantage: The adoption of GIC for depth map encoding is central to GPCGC's low-complexity claim. GIC's reliance on VQ and a simplified structure makes it significantly faster than HEVC-Intra, which is used by V-PCC.
  • RDO for Flexible Rate Control: Implementing RDO for octree splitting provides a more principled way to control the rate-distortion trade-off compared to the fixed loss functions often found in end-to-end deep learning models.

Key Results

  • Competitive Coding Performance: GPCGC achieves coding performance comparable to MPEG's V-PCC, as shown by the Rate-Distortion (RD) curves in Figure 4.
  • Significantly Lower Complexity: GPCGC exhibits substantially lower computational complexity, particularly for the decoder. The paper reports encoding FLOPs of 4G and decoding FLOPs of 64M on average, which are considerably lower than V-PCC and especially DL-based methods.
  • Smaller Model Size: Compared to DL-based methods, GPCGC has a much smaller model size (0.33 million parameters) and can handle multiple coding rates with a single model, unlike DL approaches that often require separate models for different bitrates.

Limitations and Future Work

  • Static Point Clouds: The current GPCGC method is designed for static point clouds. The authors suggest developing a "green coding solution for dynamic point clouds" as a future extension.
  • Complexity of V-PCC: While GPCGC is simpler than V-PCC, the paper notes that FLOPs only represent a fraction of V-PCC's complexity, implying that other operations might still contribute significantly.
  • DL-based Performance Gap: Although GPCGC performs comparably to V-PCC, there is still a performance gap compared to the state-of-the-art DL methods like SparsePCGC and PCGCv2, though GPCGC excels in complexity.

Summary updated

BibTeX
@inproceedings{zhou2023gpcgc,
  author = {Qingyang Zhou and Shan Liu and C.-C. Jay Kuo},
  title = {{GPCGC: A Green Point Cloud Geometry Coding Method}},
  booktitle = {ICIP},
  year = {2023},
  doi = {10.1109/icip49359.2023.10223048},
}
Figure