Slide Deformable Transformer for High-Precision LiDAR Point Cloud Compression.

Area of Science:

• Computer Vision
• Machine Learning
• Data Compression

Background:

• Dynamic LiDAR point cloud compression is crucial for reducing storage and transmission costs.
• Existing methods using Vision Transformers (ViTs) struggle with feature misalignment and inefficient attention for local motions.
• High-precision LiDAR data often loses precision due to range data quantization.

Purpose of the Study:

• To propose a novel framework for high-precision dynamic LiDAR point cloud compression.
• To address limitations of ViTs in handling cross-frame displacement and local motion.
• To preserve precision in high-accuracy LiDAR sequences.

Main Methods:

• Introduced a Slide Deformable Transformer (SDT) layer with local sliding window attention.
• Integrated deformable convolution into cross-frame attention for adaptive motion sampling.
• Developed a Radix-Decomposition Multi-Channel Quantizer (RDMCQ) for precise range data handling.

Main Results:

• The proposed SDT-PCC framework achieves high efficiency in dynamic point cloud compression.
• SDT-PCC produces temporally-coherent, accurate, and stable point cloud reconstructions.
• Experimental results on the SemanticKITTI dataset validate the framework's effectiveness.

Conclusions:

• The SDT-PCC framework offers a significant advancement in dynamic LiDAR point cloud compression.
• The proposed methods effectively mitigate feature misalignment and precision loss issues.
• This work contributes to more efficient and accurate processing of dynamic LiDAR data.