Post-Processing Geometry Enhancement for G-PCC Compressed LiDAR via Cylindrical Densification.

Area of Science:

• Computer Vision
• Signal Processing
• LiDAR Technology

Background:

• Geometry-based point cloud compression algorithms efficiently transmit sparse LiDAR data.
• Low-bitrate compression introduces severe geometry artifacts, including point reduction and coordinate offset.
• Existing methods lack post-processing solutions for direct geometry enhancement of compressed LiDAR point clouds.

Purpose of the Study:

• To develop a novel post-processing method for enhancing the geometry quality of compressed LiDAR point clouds (CLGE).
• To address geometry compression artifacts caused by low-bitrate compression modes.
• To improve the performance of downstream tasks, such as object detection, affected by compression distortion.

Main Methods:

• A two-branch approach: cylindrical densification using multi-scale sparse convolutions and adaptive refinement with MLP layers.
• Cylindrical densification employs large asymmetric sparse convolution kernels for feature extraction and candidate point generation.
• Adaptive refinement utilizes neighborhood features for refining candidate points, followed by ring-based farthest point resampling for geometry distribution preservation.

Main Results:

• The proposed method effectively enhances geometry quality in compressed LiDAR point clouds (CLGE) across various artifact levels.
• Significant reduction in compression-induced geometry artifacts, including point reduction and coordinate offset.
• The enhanced point clouds alleviate performance degradation in object detection tasks.

Conclusions:

• The developed post-processing method offers a direct solution for improving CLGE geometry quality.
• The approach is versatile, extendable to upsampling, and robust to noise.
• Geometry enhancement of CLGE improves both visual fidelity and downstream task performance.