Abstract
The proliferation of LiDAR sensors in autonomous systems has necessitated efficient Point Cloud Compression (PCC) solutions to manage the massive volumes of 3D data generated. While range image-based representations have emerged as a promising alternative to traditional tree-based methods, existing frameworks suffer from inaccurate predictions due to unidirectional temporal modeling and the mismanagement of "invalid" pixels. Furthermore, handcrafted residual coding often fails to capture complex spatial redundancies. This paper presents BIRD-PCC, a novel framework that introduces bi-directional range image prediction and deep-learned residual coding. By incorporating a mask-guided refinement process, BIRD-PCC achieves a 20% bitrate reduction over state-of-the-art baselines while maintaining high reconstruction fidelity across diverse environments.
Introduction
LiDAR sensors serve as the primary perception mechanism for autonomous vehicles, yet the high-frequency data they produce creates significant bottlenecks for storage and real-time transmission. Point Cloud Compression (PCC) techniques generally utilize either tree-based structures (e.g., Octrees) or range image projections. Range image-based methods project 3D points onto a 2D spherical plane, allowing the use of video-like compression techniques. However, current range-based models are limited by three factors: a reliance on unidirectional (past-only) prediction, the inability to distinguish between valid data and "no-return" areas (invalid pixels), and the use of generic, handcrafted residual compression algorithms like bzip2.
The BIRD-PCC Framework
BIRD-PCC (Bi-directional Range image-based Deep LiDAR PCC) addresses these limitations through a three-pronged architectural approach designed to maximize both temporal and spatial redundancy exploitation.
1. Bi-directional Prediction
Unlike conventional methods that predict the current frame ($x_t$) solely from the previous frame ($x_{t-1}$), BIRD-PCC utilizes a bi-directional scheme. Using a U-Net architecture, the model interpolates the current frame by leveraging context from both the preceding ($x_{t-1}$) and the subsequent ($x_{t+1}$) decoded frames. This approach provides the network with a more complete temporal context, leading to significantly more accurate initial estimates of the point cloud geometry.
2. Mask-Guided Refinement
LiDAR scans frequently contain "invalid" pixels where the laser pulse does not return (e.g., when pointed at the sky or distant voids). Previous models often treated these as zero-value data or noise, which blurs the resulting prediction. BIRD-PCC extracts a binary mask identifying valid versus invalid pixels. This mask is transmitted as a small piece of losslessly compressed side information. The prediction network uses this mask as a filter, ensuring the model focuses its computational resources on valid data points and preventing the introduction of artifacts in empty regions.
3. Deep-Learned Residual Coding
After the bi-directional prediction is generated, the remaining error—the residual—must be compressed. Rather than relying on standard quantization or generic compression tools, BIRD-PCC employs a Variational Autoencoder (VAE) with a hyperprior model. This neural network is specifically optimized to recognize the unique spatial structures inherent in LiDAR residuals, which differ significantly from standard RGB video residuals. This learned approach captures spatial redundancies that handcrafted methods overlook.
Experimental Evaluation
Datasets and Methodology
The framework was trained and validated using two major benchmarks:
- SemanticKITTI: Utilized for training the prediction and residual coding networks, comprising approximately 52,500 frames.
- KITTI-360: Used as an unseen test set to evaluate the model's ability to generalize to new urban environments and sensor configurations.
Performance and Key Results
BIRD-PCC demonstrates superior performance across several metrics compared to industry standards like G-PCC and state-of-the-art range-based methods like R-PCC:
- Bitrate Efficiency: BIRD-PCC achieves approximately a 20% reduction in bitrate compared to baseline range image methods at equivalent distortion levels. In the residual coding stage alone, the deep-learned approach provides a 40% saving over traditional methods.
- Prediction Accuracy: In ablation studies, the bi-directional approach combined with mask-guided side information achieved a prediction accuracy of 63.6% (at a 0.1 threshold), a substantial improvement over the 49.2% accuracy observed with simple frame averaging.
- Density Preservation: A critical advantage of this range-based approach is its ability to maintain the natural distribution and density of the point cloud. Unlike tree-based methods that can cause "voxelization" (a blocky appearance), BIRD-PCC preserves the fine-grained geometry necessary for downstream tasks like object detection and semantic segmentation.
Conclusion and Future Work
BIRD-PCC represents a significant step forward in LiDAR data management, proving that bi-directional modeling and mask-aware neural networks can drastically improve compression efficiency. While the current framework focuses on a small Group of Pictures (GOP) and geometry compression, future research will aim to scale the architecture for longer sequences and incorporate attribute compression (e.g., intensity and semantic labels). Additionally, optimizing the computational overhead for real-time deployment on edge hardware remains a priority for the next iteration of the framework.
Summary updated