Point Cloud Attribute Coding With Bottom-Up Octree Features

The patent inverts the traditional top-down probability estimation approach by implementing a bottom-up method. Instead of estimating probabilities from parent level information, the system first extracts features from finer child level voxels, then aggregates these features upward to improve probability estimation accuracy at each octree level, thereby resolving the contradiction between accuracy and complexity

The patent introduces a new dimensional approach by extracting features from multiple child level voxels and aggregating them through neural network operations. This multi-dimensional feature aggregation from finer levels provides richer information for probability estimation without following the traditional single-direction parent-to-child approach, thus improving accuracy while managing complexity

The patent segments the point cloud data into octree voxels at multiple levels, allowing feature extraction to be performed on smaller, manageable child level segments. This segmentation enables efficient processing of finer level details by breaking down the complex task into hierarchical sub-tasks, improving compression efficiency while controlling processing complexity through divide-and-conquer

The patent implements a nested hierarchical structure where child level voxels are nested within parent level voxels. Feature extraction and aggregation operate within this nested framework, allowing efficient utilization of finer level details by processing nested sub-structures first and then aggregating results upward, thereby improving compression efficiency while managing complexity through hierarchical organization