Inter-Block Intra Prediction for Parallel Point Cloud Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional RAHT transform in point cloud compression cannot be parallelized, hindering real-time applications due to its requirement of processing the entire point cloud frame as a whole, preventing independent encoding of sub-blocks of the voxel grid.
Innovation Solution
Divide the point cloud into blocks for encoding/decoding, employing a global step to encode the average color of each block and a local step to encode attributes, allowing parallel execution of intra prediction using neighboring blocks for improved efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional RAHT transform processes the entire point cloud frame as a whole, then encoding accuracy is maintained, but processing speed deteriorates and real-time applications are hindered
Solution Approach 1:
The patent divides the point cloud frame into multiple independent sub-blocks (e.g., 8x8 or 16x16 voxel grids) that can be encoded and decoded in parallel. Each sub-block is processed independently with its own intra-prediction referencing neighboring sub-blocks, enabling parallelization while maintaining overall encoding quality through the global step that processes average colors at block boundaries.
2Productivity
If the entire point cloud frame is processed sequentially, then encoding accuracy is preserved, but real-time processing capability is lost
Solution Approach 1:
The point cloud attribute data is segmented into multiple independently encodable sub-blocks arranged in a grid pattern. Each sub-block can be processed simultaneously by different processing units, transforming sequential processing into parallel execution. The segmentation enables real-time processing by reducing the critical path length while maintaining compression efficiency through coordinated boundary handling.
Solution Approach 2:
The global step performs preliminary encoding of average color values for each block before the local step processes detailed attributes. This preliminary action establishes reference data at block boundaries that enables subsequent parallel local encoding without requiring global coordination during the actual attribute encoding phase, reducing processing time while preserving accuracy.
3Productivity
If independent sub-block encoding is implemented, then parallel processing is enabled, but reference data availability across block boundaries becomes problematic
Solution Approach 1:
The patent introduces a global step as an intermediary that processes average color values at block boundaries and generates reference data that is made available to all local steps. This intermediary layer ensures that neighboring blocks have access to necessary reference information for accurate intra-prediction, preventing information loss while enabling parallel processing of sub-blocks.
Solution Approach 2:
Reference data for block boundaries is prepared in advance by the global step before the parallel local encoding begins. This preliminary preparation of reference information ensures that all sub-blocks have access to necessary data for accurate prediction without requiring inter-communication during parallel execution, maintaining information availability while enabling independence of parallel processes.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In a point cloud codec, attributes correspond to nodes in a tree structure. The tree is partitioned into subtrees, each subtree having a respective root node. The root nodes may all be at a layer Lglobal. Attributes of the subtree root nodes may be coded using intra coding, with residuals coded using a region-adaptive hierarchical transform (RAHT). To allow for parallel decoding, nodes of the subtrees may be intra predicted using neighboring nodes from different subtrees as reference nodes even if they have not yet been reconstructed, e.g. by temporarily using the attributes of the subtree root node as the attribute of one or more of the reference nodes for purposes of the intra prediction. Those temporary attributes may be replaced with the actual attributes once the nodes in the subtree have been reconstructed (e.g. using a local RAHT for each subtree).