Introduction
Point clouds are essential for immersive technologies like VR/AR and autonomous systems, but their massive data size poses a significant storage and transmission challenge. While geometry compression is well-handled by octree-based methods, attribute compression (color, reflectance) often struggles to balance computational efficiency with spatial awareness. Current standards like the Region Adaptive Haar Transform (RAHT) are efficient but limited, as they only consider two nodes at a time, missing complex spatial correlations. Conversely, global Graph Fourier Transforms (GFT) capture these relationships but are computationally prohibitive for large datasets. This paper introduces the Successive Subspace Graph Transform (SSGT) to provide a computationally feasible, spatially aware middle ground.
The SSGT Framework
The SSGT method applies the principles of Successive Subspace Learning (SSL) to a hierarchical graph structure derived from an octree. The process begins by partitioning the 3D space into nested cubes. At each level, a weighted graph is constructed where nodes represent subspaces.
A key innovation is the inclusion of self-loops in the graph design. The weights of these loops are proportional to the number of points within a subspace, making the transform "density-aware." The transform uses a Normalized Laplacian GFT to decompose color signals into low-frequency (DC) and high-frequency (AC) components. Starting from the leaf nodes, the transform is applied successively; DC coefficients are passed up to parent nodes, allowing the system to scale to millions of points by processing small, local sub-graphs at each stage.
Key Insights and Advantages
- Generalization of RAHT: Mathematically, RAHT is a special case of SSGT limited to a two-vertex graph. By expanding the graph size (e.g., to $4 \times 4 \times 4$ voxel blocks), SSGT captures more intricate spatial patterns.
- Orientation Robustness: Unlike axis-aligned transforms, SSGT’s distance-based weighted graphs reduce "axis bias," making the compression more consistent regardless of the point cloud's orientation.
- Scalability: By processing sub-graphs successively from the bottom up, the method avoids the $O(N^3)$ complexity of global GFTs while maintaining superior energy compaction compared to local Haar transforms.
Experimental Results
The SSGT method was evaluated using the 8i dynamic point cloud datasets (e.g., "longdress", "soldier"). Performance was measured using Rate-Distortion (R-D) curves, comparing Y-PSNR against bits per point (bpp).
As shown in the R-D curves, SSGT consistently outperforms RAHT across all test sequences. For the "longdress" sequence at depth 10, SSGT achieved a Y-PSNR of 35.34 dB at 1.63 bpp, whereas RAHT achieved a lower 34.67 dB at a higher 1.72 bpp. Furthermore, larger sub-graph constructions (e.g., $4 \times 4 \times 4$ voxels) yielded better compression performance than smaller ones ($2 \times 2 \times 2$), confirming that capturing more neighbors in the graph transform leads to better energy compaction.
Conclusion
SSGT offers a scalable and spatially aware alternative for point cloud attribute compression. While it introduces more computational complexity than RAHT due to the need for eigenvalue decomposition in sub-graphs, the gains in compression efficiency and visual fidelity are significant. Future work will focus on optimizing graph designs for non-uniform point distributions and exploring more sophisticated geometry representations beyond simple cubic subspaces.
Summary updated