Sparse Tensor-Based Multiscale Representation for Point Cloud Geometry Compression.

Area of Science:

• Computer Vision
• Data Compression
• Geometric Processing

Background:

• Point Cloud Geometry (PCG) compression is crucial for efficient data handling.
• Existing methods often struggle with balancing compression efficiency and computational complexity.
• Multiscale and sparse representations offer potential for improved PCG compression.

Purpose of the Study:

• To develop a unified and efficient compression method for Point Cloud Geometry (PCG).
• To reduce computational complexity by focusing on sparsely distributed voxels.
• To enhance compression efficiency by exploiting multiscale correlations.

Main Methods:

• Developed SparsePCGC, a multiscale sparse tensor-based voxelized PCG compression method.
• Utilized Sparse Convolution-based Neural Network (SparseCNN) for characterizing spatial correlations.
• Implemented SparseCNN-based Occupancy Probability Approximation (SOPA) for probability estimation.
• Incorporated SparseCNN based Local Neighborhood Embedding (SLNE) for improved feature attributes.

Main Results:

• Achieved state-of-the-art performance in both lossless and lossy compression modes.
• Demonstrated superior results on dense (8iVFB, Owlii, MUVB) and sparse (KITTI, Ford) datasets.
• Outperformed standardized MPEG G-PCC and other learning-based compression schemes.
• Maintained low computational complexity, making it suitable for practical applications.

Conclusions:

• SparsePCGC offers a unified, efficient, and low-complexity solution for PCG compression.
• The multiscale sparse tensor approach effectively exploits spatial and cross-scale correlations.
• The method shows significant advantages over existing techniques for diverse point cloud data.