ECM-OPCC: Efficient Context Model for Octree-based Point Cloud Compression

Yiqi Jin, Ziyu Zhu, Tongda Xu, Yuhuan Lin, Yan Wang

TL;DR

This paper introduces ECM-OPCC, a novel deep learning framework for efficient octree-based point cloud compression. It tackles the trade-off between compression performance and decoding speed by proposing a multi-group coding strategy and a dual transformer architecture, achieving state-of-the-art compression while drastically reducing decoding time.

What Problem Does This Paper Address?

Point cloud compression is crucial for handling the massive datasets generated by 3D sensors. Octree-based methods, which recursively subdivide space, are popular for their ability to represent sparse data. However, existing deep learning approaches for octree compression often rely on complex autoregressive context models that capture dependencies between nodes (ancestors and siblings). While this leads to good compression ratios, it results in extremely high decoding complexity, making real-world deployment challenging. For instance, the state-of-the-art OctAttention method can take over 900 seconds to decode a single point cloud. This paper aims to bridge this gap by developing a context model that is both sufficient for high compression and efficient for fast decoding.

Methods

The core of ECM-OPCC lies in its novel segment-constrained multi-group coding strategy and a dual transformer architecture.

  • Multi-Group Coding Strategy: Instead of a fully sequential autoregressive approach, the authors divide the nodes within each octree layer into "context segments." These segments are then further divided into "groups." Nodes within a group are decoded in parallel, leveraging context from previously decoded groups within the same segment and from previous layers. This parallelization significantly speeds up decoding.
  • Dual Transformer Architecture: This architecture comprises two branches:
  • Level-Parallel Branch: This branch captures context from ancestor nodes (previous layers) and is designed to allow parallel decoding within a layer, aligning with Eq. 5 in the paper.
  • Group-Parallel Branch: This branch focuses on exploiting context from sibling nodes within the same layer. It uses a multi-group mask to ensure that only already decoded sibling information is used as context for current nodes, enabling parallel decoding within groups.
  • Random Masking Pre-training: To enhance the model's ability to predict node occupancy from context, a BERT-inspired random masking pre-training strategy is employed. Input occupancy codes are randomly masked, and the model learns to predict these masked values using unmasked context.

Datasets

The authors evaluate their method on two types of datasets:

  • LiDAR Dataset: SemanticKITTI [1] is used for lossy compression experiments.
  • Object Datasets: MPEG 8i Voxelized Full Bodies [17] and Microsoft Voxelized Upper Bodies (MVUB) [4] are used for lossless compression experiments. These datasets are commonly used for lossless compression tasks due to their fixed precision.

No new benchmarks were introduced.

Key Insights and Ideas

  • Parallelism is Key to Decoding Speed: The most significant insight is that by strategically dividing the autoregressive context into parallelizable segments and groups, the decoding complexity can be drastically reduced without sacrificing compression performance. This is a crucial step towards practical deep learning-based point cloud compression.
  • Dual Transformers for Complementary Context: The dual transformer architecture effectively captures both hierarchical (ancestor) and spatial (sibling) context, demonstrating that a combined approach is more powerful than relying on one type of context alone.
  • Pre-training Matters: The random masking pre-training strategy significantly boosts the model's performance, highlighting the importance of robust context modeling for accurate occupancy prediction.

Key Results

  • State-of-the-Art Compression: ECM-OPCC achieves state-of-the-art performance for both lossy and lossless point cloud compression across various datasets.
  • Dramatic Decoding Speedup: The multi-group coding strategy leads to a remarkable reduction in decoding time, saving up to 98% of decoding time compared to the previous octree-based SOTA method (OctAttention). For example, OctAttention takes ~948s, while ECM-OPCC takes ~19.5s.
  • Competitive Encoding Time: While the dual transformer architecture increases model size and computation, the overall encoding time remains competitive, and the decoding speedup far outweighs this.
  • Lossy Compression: On the SemanticKITTI dataset, ECM-OPCC outperforms G-PCC by ~31% on average across distortion levels and achieves better bitrates than OctAttention at high and low bitrates.
  • Lossless Compression: On MPEG 8i and MVUB datasets, ECM-OPCC saves ~42.4% and ~33.6% bpp over G-PCC, respectively. With pre-training, it outperforms OctAttention by ~6.3% and ~1.3%.

Limitations and Future Work

  • Model Size and FLOPs: The paper acknowledges that their model has a larger parameter count and higher FLOPs compared to OctAttention due to the dual-channel structure. However, they argue that the efficiency gains from their coding strategy compensate for this.
  • Hardware Parallelism Assumption: The theoretical speedup analysis assumes fully parallelizable hardware. While practical hardware offers parallelism, the exact speedup might vary.
  • Future Work: The authors suggest exploring further optimizations for the dual transformer architecture and investigating more advanced pre-training strategies. They also imply that their framework can be extended to other point cloud representations beyond octrees.

Summary updated

Abstract

Recently, deep learning methods have shown promising results in point cloud compression. However, previous octree-based approaches either lack sufficient context or have high decoding complexity (e.g. > 900s). To address this problem, we propose a sufficient yet efficient context model and design an efficient deep learning codec for point clouds. Specifically, we first propose a segment-constrained multi-group coding strategy to exploit the autoregressive context while maintaining decoding efficiency. Then, we propose a dual transformer architecture to utilize the dependency of current node on its ancestors and siblings. We also propose a random-masking pre-train method to enhance our model. Experimental results show that our approach achieves state-of-the-art performance for both lossy and lossless point cloud compression, and saves a significant amount of decoding time compared with previous octree-based SOTA compression methods.
BibTeX
@inproceedings{jin2022ecm,
  author = {Yiqi Jin and Ziyu Zhu and Tongda Xu and Yuhuan Lin and Yan Wang},
  title = {{ECM-OPCC: Efficient Context Model for Octree-based Point Cloud Compression}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=VAA3NeklEq},
}
Figure