PaperlayerView exact revision on arXiv

Source-linked AI summary

VoxelContext-Net: An Octree based Framework for Point Cloud Compression

Zizheng Que, Guo Lu, Dong Xu

arXiv:2105.02158v1cs.CVeess.IV

TL;DR

Point cloud compression must handle sparse, orderless 3D data and limitations in existing tree, voxel, and point-based methods. VoxelContext-Net combines octree organization with local and temporal voxel context, lossless entropy coding, and decoder-side coordinate refinement. It achieves state-of-the-art performance across static and dynamic benchmark datasets, with reported bitrate savings over G-PCC on ScanNet and Semantic KITTI.

  • Problem

    Existing point cloud compression methods face efficiency, sparsity, end-to-end optimization, spatial-context, and octree-distortion limitations.

  • Method

    VoxelContext-Net uses local voxel context in an octree framework for lossless non-leaf-symbol entropy coding, temporal context for dynamic data, and decoder-side coordinate refinement.

  • Results

    43.66% bitrate saving on ScanNet and 31.15% on Semantic KITTI are reported against G-PCC for static compression.

  • Takeaways & Limitations

    The framework provides a unified learning-based approach for static and dynamic point cloud geometry compression with state-of-the-art performance on two benchmark datasets.

  • Takeaways & Limitations

    Traditional tree-based methods remain constrained by hand-crafted techniques that cannot be optimized end-to-end using large-scale data.

Abstract

from arXiv · show

In this paper, we propose a two-stage deep learning framework called VoxelContext-Net for both static and dynamic point cloud compression. Taking advantages of both octree based methods and voxel based schemes, our approach employs the voxel context to compress the octree structured data. Specifically, we first extract the local voxel representation that encodes the spatial neighbouring context information for each node in the constructed octree. Then, in the entropy coding stage, we propose a voxel context based deep entropy model to compress the symbols of non-leaf nodes in a lossless way. Furthermore, for dynamic point cloud compression, we additionally introduce the local voxel representations from the temporal neighbouring point clouds to exploit temporal dependency. More importantly, to alleviate the distortion from the octree construction procedure, we propose a voxel context based 3D coordinate refinement method to produce more accurate reconstructed point cloud at the decoder side, which is applicable to both static and dynamic point cloud compression. The comprehensive experiments on both static and dynamic point cloud benchmark datasets(e.g., ScanNet and Semantic KITTI) clearly demonstrate the effectiveness of our newly proposed method VoxelContext-Net for 3D point cloud geometry compression.

1. Introduction

VoxelContext-Net addresses point cloud compression challenges by combining octree organization with voxel-context modeling. It supports static and dynamic compression, lossless entropy coding, decoder-side coordinate refinement, and strong benchmark performance.

  • Point cloud compression is challenging because 3D points are sparse and orderless, while voxel methods can incur high computational complexity.
  • VoxelContext-Net exploits voxel context within an octree framework to combine efficient data organization with spatial modeling.
  • The method losslessly compresses non-leaf octree symbols using local voxel representations of neighboring nodes at the same depth.
  • For dynamic point clouds, the approach additionally uses voxel representations from temporally neighboring point clouds.
  • Experiments on large-scale static and dynamic datasets show state-of-the-art compression performance and outperform both hand-crafted and learning-based methods.
  • A voxel-context coordinate refinement module predicts more accurate decoder-side coordinates for leaf nodes.

2. Related Work

Prior point cloud compression methods use hand-crafted tree techniques, point-based networks, or voxelized representations with efficiency limitations. VoxelContext-Net addresses these limitations by modeling neighboring octree nodes, refining coordinates, and extending context across frames.

  • Traditional tree-based codecs rely on hand-crafted techniques and cannot be optimized end-to-end using large-scale data.
  • Point-based methods may require substantial memory and computational cost, while voxel-based methods ignore point-cloud sparsity.
  • OctSqueeze uses ancestor-node context but ignores spatially neighboring nodes at the same depth and octree-construction distortion.
  • OctSqueeze is designed for static compression, and a concurrent dynamic method retains similar limitations.
  • VoxelContext-Net models neighboring nodes with local voxel representations, refines decoder-side coordinates, and uses neighboring-frame context for dynamic compression.

3. Methodology

The proposed method organizes point clouds with octrees, compresses non-leaf symbols through a voxel-context entropy model, and refines reconstructed coordinates at the decoder. The architecture is introduced for static compression and extended to dynamic point clouds.

  • Dynamic extension: The architecture is presented for static point cloud compression and then extended to dynamic point clouds.
  • Static compression: The framework first organizes the input point cloud using an octree and targets lossless coding of non-leaf-node symbols.
  • Static compression: A voxel-context deep entropy model predicts symbol probability distributions to improve compression performance.
  • Static compression: A local voxel-context coordinate refinement module compensates for octree-construction distortion and produces more accurate reconstructed points.

3.2. Octree Construction

The octree recursively partitions 3D space into eight cubes until a maximum depth, representing each non-leaf node with an occupancy symbol. Because leaf centers are quantized cube centers, reconstruction can deviate from original coordinates.

  • An octree recursively partitions 3D space into eight equal cubes for each non-empty region until reaching the maximum depth level.
  • Each node’s 3D coordinate denotes its cube center, and each non-leaf node uses an 8-bit symbol for child occupancy.
  • Octree depth affects reconstruction quality because a leaf-node cube center may not match the original point coordinate.
  • For example, the original coordinate (0.6, 0.7, 0.7) is quantized to the leaf-center coordinate (0.625, 0.625, 0.625), introducing distortion.

3.3. Local Voxel Context in Octree

The method converts octree structure into binary voxel representations at each depth, then extracts local voxel context around each node to capture same-level neighboring occupancy.

  • Figure 3 traces the process from the raw point cloud and octree to the depth-2 voxel grid and its detailed binary representation.The purple region identifies the local context centered on the current node.
  • The point cloud is represented as a binary voxel grid whose entries indicate whether points exist in the corresponding spatial cubes.At depth k, the grid has shape 2^k × 2^k × 2^k.
  • For each node, the local voxel representation Vi ∈ R^M×M×M records neighboring-node distribution at the same octree depth.This local context is used as prior information for compression.

3.4. Our Deep Entropy Model

The deep entropy model estimates occupancy-symbol probabilities from each node’s local voxel context and node features, enabling lossless compression of non-leaf octree symbols.

  • Each non-leaf octree node contributes an 8-bit occupancy symbol describing which of its eight children exist.The sequence s collects these symbols across all non-leaf nodes.
  • The estimated distribution is trained with cross-entropy against the unknown occupancy distribution, supporting lossless symbol compression.The learned model approximates the actual distribution because the practical distribution P is unavailable.
  • The model assumes a node’s occupancy symbol depends on its local voxel context Vi and node feature ci, including its coordinate and depth.This assumption simplifies the complex dependence among decoded and neighboring nodes.
  • A CNN extracts spatial context features from Vi, an MLP fuses them with ci, and a softmax predicts qs(si|Vi, ci).The architecture uses convolutional processing followed by feature fusion and probability prediction.

3.5. Our Coordinate Refinement Method

The coordinate refinement method corrects leaf-node coordinates after octree decoding by predicting context-dependent offsets from neighboring voxels.

  • Octree decoding yields a coordinate for each leaf node, but the refinement module predicts a more accurate output coordinate.The goal is to reduce distortion introduced during octree construction.
  • For leaf node ni, the refined coordinate is obtained by adding a learned offset R(Vi) from its local voxel context to the decoded coordinate.R(·) is the learnable coordinate-refinement function.
  • The module converts the decoded octree into a global binary voxel representation and extracts each leaf’s local context.CNNs encode 3D context, followed by fully connected layers that predict the coordinate offset.

3.6. Training Strategy

The entropy model and coordinate refinement module are trained separately, using occupancy prediction loss and coordinate reconstruction loss, respectively.

  • The deep entropy model and coordinate refinement module are optimized independently rather than in a joint training procedure.Separate training is specified for the two components.
  • The entropy model is trained from predicted occupancy distributions qs(si|Vi, ci) at non-leaf nodes.Its loss is based on the predicted probabilities for the octree occupancy symbols.
  • For dynamic point clouds, four voxel representations from decoded layers and neighboring frames provide context for estimating current-node probabilities.The available temporal representations are aligned using sensor pose information.
  • The coordinate refinement module uses mean squared error to train precise coordinates for reconstructed points.The target is the ground-truth coordinate before octree construction.

3.7. Dynamic Point Cloud Compression

For dynamic point cloud compression, VoxelContext-Net aligns neighboring frames and fuses their voxel contexts to model temporal dependencies. It then refines reconstructed coordinates using the current frame’s voxel representation.

  • Dynamic compression aligns P_t, P_t−1, and P_t+1 into a common coordinate system using sensor pose information.
  • The entropy model extracts features from four voxel representations spanning the current and neighboring point clouds, then concatenates them for fusion.
  • Coordinate refinement uses only the current point cloud’s voxel representation because it is sufficient to produce promising results.

4. Experiments

Experiments on ScanNet and Semantic KITTI evaluate static and dynamic compression, ablations, computational trade-offs, and downstream task performance. VoxelContext-Net consistently improves compression results, while temporal context benefits dynamic compression and coordinate refinement improves the full model.

  • Experimental Details: The experiments use ScanNet and Semantic KITTI, with Semantic KITTI supporting both static and dynamic compression evaluations.
  • Static Point Cloud Compression: VoxelContext-Net saves 43.66% bitrate on ScanNet and 31.15% on Semantic KITTI relative to G-PCC, exceeding OctSqueeze’s 15.00% and 2.13% savings.
  • Static Point Cloud Compression: At 4 bpp on ScanNet, VoxelContext-Net improves PSNR by more than 2 dB over OctSqueeze.
  • Static Point Cloud Compression: On Semantic KITTI, VoxelContext-Net reaches 55.51 dB PSNR at 0.36 bpp, while OctSqueeze reaches 54.35 dB at 0.52 bpp.
  • Dynamic Point Cloud Compression: Dynamic VoxelContext-Net saves 38.10% bitrate against G-PCC, compared with 5.01% for OctSqueeze, and saves an additional 11.11% over the static version.
  • Dynamic Point Cloud Compression: Estimated-pose dynamic compression saves 41.77% bitrate and still outperforms the existing baseline methods.
  • Ablation Study and Analysis: Local voxel context saves an additional 14.81% BDBR over OctSqueeze, while coordinate refinement saves a further 21.50% BDBR.
  • Ablation Study and Analysis: Increasing voxel resolution improves compression, but larger representations increase computational complexity; 9 × 9 × 9 provides the selected trade-off.

5. Conclusion

The paper concludes that VoxelContext-Net combines local voxel context, lossless octree entropy coding, and coordinate refinement for static and dynamic point cloud geometry compression. Experiments on two benchmark datasets report state-of-the-art compression performance.

  • VoxelContext-Net uses local voxel representations within an octree-structured point cloud to support learning-based geometry compression.
  • Its deep entropy model losslessly compresses octree-node symbols, while coordinate refinement reconstructs higher-quality point clouds at the decoder.
  • The framework applies to both static and dynamic point cloud compression and achieves state-of-the-art performance on two benchmark datasets.
Loading 2105.02158v1…