BIRD-PCC: Bi-Directional Range Image-Based Deep Lidar Point Cloud Compression

Chia-Sheng Liu, Jia-Fong Yeh, Hao Hsu, Hung-Ting Su, Ming-Sui Lee, Winston H. Hsu

TL;DR

BIRD-PCC is a novel LiDAR point cloud compression framework that leverages bi-directional range image prediction and deep-learned residual coding. By incorporating a "mask-guided" approach that accounts for invalid pixels (no-return areas), it achieves a 20% bitrate reduction over state-of-the-art baselines while maintaining superior reconstruction quality.

What Problem Does This Paper Address?

LiDAR sensors are the "eyes" of autonomous vehicles, but they generate massive amounts of 3D data that are costly to store and transmit. Existing Point Cloud Compression (PCC) methods generally fall into two camps: tree-based (like Octrees) and range image-based (projecting 3D points onto a 2D spherical plane).

The authors identify three major flaws in current range image-based methods:

  1. Unidirectional Prediction: Most methods only look at past frames to predict the current one, missing out on the context provided by future frames.
  2. Ignoring "Invalid" Pixels: LiDAR often receives no signal return (e.g., pointing at the sky), resulting in "invalid" pixels. Previous models treat these as zeros or noise, leading to blurry and inaccurate predictions.
  3. Inefficient Residual Coding: After making a prediction, the "leftover" error (residual) is usually compressed using handcrafted, lossless methods like bzip2. These methods fail to capture complex spatial patterns, leaving a lot of compression potential on the table.

Methods

BIRD-PCC (Bi-directional Range image-based Deep LiDAR PCC) introduces a three-pronged strategy to solve these issues:

  • Bi-directional Prediction: Instead of just looking backward, the model uses a U-Net architecture to interpolate the current frame ($x_t$) using both the previous ($x_{t-1}$) and the next ($x_{t+1}$) decoded frames.
  • Mask-Guided Refinement: The system extracts a binary mask that identifies which pixels are valid. This mask is sent as a small piece of "side information" (compressed losslessly). The prediction network uses this mask as a filter, ensuring the model doesn't waste effort trying to predict values where no data exists, which significantly reduces "noise" in the reconstructed point cloud.
  • Deep-Learned Residual Coding: Rather than using standard quantization and generic compression, the authors employ a Variational Autoencoder (VAE) with a hyperprior model. This neural network is specifically trained to recognize and compress the spatial redundancies found in LiDAR residuals, which are very different from the residuals found in standard color video.

Datasets

The researchers used two primary datasets to train and validate their approach:

  • SemanticKITTI: Used for training the prediction and residual coding networks (approx. 52.5k frames total).
  • KITTI-360: Used as an unseen, larger test set to evaluate how well the model generalizes to new environments. This is a crucial step to ensure the AI hasn't just "memorized" one specific city's layout.

Key Insights and Ideas

  • The "Invalid" Advantage: One of the most clever insights is treating the absence of data as a feature. By explicitly telling the network where the "holes" in the data are via the binary mask, the prediction accuracy (measured by L1 and RMSE) improved drastically.
  • Neural Residuals: The paper demonstrates that LiDAR residuals have distinct spatial structures. A deep-learned encoder can "see" these patterns much better than a generic algorithm like bzip2, leading to a 40% bitrate saving in the residual coding stage alone.
  • Density Preservation: Unlike tree-based methods (like G-PCC) which can cause "voxelization" (making the world look like it's made of blocks), the range image approach maintains the natural distribution and density of the points, which is vital for downstream tasks like object detection.

Key Results

  • Efficiency: BIRD-PCC costs nearly 20% less bitrate than baseline range image-based methods at the same distortion level.
  • Performance: It outperforms the industry-standard G-PCC at high bitrates and significantly beats R-PCC (a leading range-based method) across low bitrates.
  • Accuracy: In ablation studies, the bi-directional approach with side information (the mask) achieved an accuracy of 63.6% (at a 0.1 threshold), compared to only 49.2% for simple frame averaging.

Limitations and Future Work

While BIRD-PCC shows significant gains, there are areas for further exploration:

  • GOP Scaling: The current work focuses on a small "Group of Pictures" (GOP), specifically predicting one frame between two intra-frames ($k=1$). Future work could explore how this scales to longer sequences of inter-frames.
  • Computational Overhead: Deep learning-based compression is often more computationally intensive than handcrafted methods. The paper does not provide a detailed breakdown of the latency or power consumption required for real-time deployment on vehicle hardware.
  • Attribute Compression: The study focuses on geometry (coordinates). Future iterations could incorporate point attributes like intensity or semantic labels into the compression pipeline.

Summary updated

Abstract

The large amount of data collected by LiDAR sensors brings the issue of LiDAR point cloud compression (PCC). Previous works on LiDAR PCC have used range image representations and followed the predictive coding paradigm to create a basic prototype of a coding framework. However, their prediction methods give an inaccurate result due to the negligence of invalid pixels in range images and the omission of future frames in the time step. Moreover, their handcrafted design of residual coding methods could not fully exploit spatial redundancy. To remedy this, we propose a coding framework BIRD-PCC. Our prediction module is aware of the coordinates of invalid pixels in range images and takes a bidirectional scheme. Also, we introduce a deep-learned residual coding module that can further exploit spatial redundancy within a residual frame. Experiments conducted on SemanticKITTI and KITTI-360 datasets show that BIRD-PCC outperforms other methods in most bitrate conditions and generalizes well to unseen environments.
BibTeX
@inproceedings{liu2023bird,
  author = {Chia-Sheng Liu and Jia-Fong Yeh and Hao Hsu and Hung-Ting Su and Ming-Sui Lee and Winston H. Hsu},
  title = {{BIRD-PCC: Bi-Directional Range Image-Based Deep Lidar Point Cloud Compression}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10095458},
}

Loading related content...

Figure