Towards Point Cloud Compression for Machine Perception: A Simple and Strong Baseline by Learning the Octree Depth Level Predictor

Lei Liu, Zhihao Hu, Zhenghao Chen

TL;DR

This paper introduces PCCMP-Net, a point cloud compression framework that adaptively adjusts bit-rates based on the needs of machine vision tasks. By learning an Octree Depth Level Predictor, the system sends only the necessary amount of data for tasks like classification or detection, saving up to 45% in bit-rate without sacrificing human-viewing quality.

What Problem Does This Paper Address?

As LiDAR and RGB-D sensors become ubiquitous in autonomous vehicles and robotics, the volume of 3D point cloud data is exploding. Traditional Point Cloud Compression (PCC) algorithms are designed for human vision, aiming to make the reconstructed 3D model look "right" to a person.

However, most point cloud data is actually consumed by machines for tasks like object detection or semantic segmentation. These "machine vision" tasks often don't require the full, high-resolution bit-stream. For example, a neural network can identify an airplane from a very sparse set of points, whereas a human might need a dense reconstruction to inspect the surface. Current compression methods waste bandwidth by sending more detail than a machine needs, or they degrade human-viewing quality to satisfy machine requirements.

Methods

The authors propose a "scalable coding" approach. Instead of one-size-fits-all compression, they partition the bit-stream so that machine tasks use a subset of the data, while the full stream remains available for human inspection.

The Core Innovation: Octree Depth Level Predictor (ODL)

Most modern PCC frameworks (like G-PCC or OctAttention) represent point clouds using octrees—a tree structure where space is recursively divided into eight sub-cubes. The deeper the tree, the higher the resolution (and the higher the bit-rate).

The authors introduced a lightweight ODL Predictor (a simple 3-layer MLP). For any given point cloud and task (e.g., "detect cars"), the ODL predicts the optimal octree depth.

  • Simple tasks/objects: The ODL selects a shallow depth, sending fewer bits.
  • Complex tasks/objects: The ODL selects a deeper level to ensure accuracy.

Making it Differentiable

Choosing a discrete depth level is normally a non-differentiable operation, which makes it hard to train using standard backpropagation. To solve this, the authors used the Gumbel-Softmax trick, allowing the model to learn the best depth-selection strategy end-to-end.

Datasets

The framework was evaluated across five diverse benchmarks to prove its generalizability:

  • ModelNet10 & ModelNet40: For 3D object classification.
  • ShapeNet: For part segmentation.
  • ScanNet: For indoor scene segmentation and detection.
  • KITTI: For outdoor autonomous driving detection and segmentation.

Key Insights and Ideas

  • Task-Specific Granularity: Not all 3D objects are created equal. A "table" is geometrically simple and can be segmented at a low octree depth. An "airplane" with thin wings requires a much deeper octree to be recognized correctly. The ODL learns these nuances automatically.
  • Plug-and-Play Compatibility: The ODL isn't a standalone compressor; it’s a "wrapper" that can be integrated into existing mainstream frameworks like VoxelContext-Net, OctAttention, and the industry-standard G-PCC.
  • Human-Machine Synergy: By using a scalable bit-stream, the system avoids the "either-or" trap. You save bandwidth on the machine side, but if a human operator needs to look at the data later, the remaining bits can be transmitted to reconstruct the full-resolution model.

Key Results

The results show that "less is more" for machine perception:

  • Bit-rate Savings: On the ModelNet40 dataset, PCCMP-Net saved approximately 45% of the bit-rate while maintaining a 91.8% classification accuracy.
  • Performance Boost: Surprisingly, using the ODL often improved machine performance compared to standard baselines. For instance, it achieved a 10% accuracy improvement at very low bit-rates (0.056 bpp) compared to VoxelContext-Net.
  • Zero Human Cost: Because the full bit-stream is preserved as the "top layer" of the hierarchy, there is no degradation in PSNR (Point-to-Point quality) for human vision tasks.

Limitations and Future Work

While the baseline is strong, the authors note a few areas for growth:

  • Encoder Power: The ODL currently assumes the encoder has enough intelligence to run the predictor. However, some low-power LiDAR sensors might struggle with even lightweight MLPs.
  • Task Specificity: Currently, the ODL is trained for specific tasks. A more universal predictor that can balance multiple simultaneous machine tasks (e.g., detecting and segmenting at once) would be a logical next step.
  • Dynamic Scenes: The current work focuses largely on static point clouds; extending this adaptive depth logic to temporal sequences (video point clouds) remains an open challenge.

Summary updated

Abstract

Point cloud compression has garnered significant interest in computer vision. However, existing algorithms primarily cater to human vision, while most point cloud data is utilized for machine vision tasks. To address this, we propose a point cloud compression framework that simultaneously handles both human and machine vision tasks. Our framework learns a scalable bit-stream, using only subsets for different machine vision tasks to save bit-rate, while employing the entire bit-stream for human vision tasks. Building on mainstream octree-based frameworks like VoxelContext-Net, OctAttention, and G-PCC, we introduce a new octree depth-level predictor. This predictor adaptively determines the optimal depth level for each octree constructed from a point cloud, controlling the bit-rate for machine vision tasks. For simpler tasks (\textit{e.g.}, classification) or objects/scenarios, we use fewer depth levels with fewer bits, saving bit-rate. Conversely, for more complex tasks (\textit{e.g}., segmentation) or objects/scenarios, we use deeper depth levels with more bits to enhance performance. Experimental results on various datasets (\textit{e.g}., ModelNet10, ModelNet40, ShapeNet, ScanNet, and KITTI) show that our point cloud compression approach improves performance for machine vision tasks without compromising human vision quality.
BibTeX
@misc{liu2024towards,
  author = {Lei Liu and Zhihao Hu and Zhenghao Chen},
  title = {{Towards Point Cloud Compression for Machine Perception: A Simple and Strong Baseline by Learning the Octree Depth Level Predictor}},
  year = {2024},
  eprint = {2406.00791},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.00791},
}
Figure