Point Cloud Attribute Coding with Parameterized Voxel Probabilities

The patent extends the probability estimation from the traditional top-down approach (using only parent/sibling nodes) to a multi-dimensional approach that incorporates neighbor nodes at the same level. This adds a new dimension of spatial correlation consideration, allowing the model to capture relationships between neighboring classes that were previously ignored, thereby improving compression performance while maintaining implementation feasibility through systematic integration into the existing octree framework

The patent employs parameterized probability distributions (such as Gaussian or Laplace distributions) to model attribute correlations between neighboring classes. Instead of using complex non-parametric methods or exhaustive enumeration, the approach uses a small set of distribution parameters (mean, variance, etc.) that can be efficiently estimated from the feature map. This parameterization strategy achieves high probability estimation accuracy while keeping computational complexity manageable, as the parameters can be computed through standard statistical operations on the voxel features