Matrix Intra Prediction Low-Rank Factorization for Video Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The use of Matrix Intra Prediction (MIP) in video encoding and decoding requires significant storage and computational resources due to large matrices, limiting its applicability and efficiency.
Innovation Solution
Replace the original MIP prediction matrices with the product of two lower-rank matrices, reducing storage and computation requirements while maintaining prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If original MIP prediction matrices are used, then prediction accuracy is maintained, but storage and computational resources are significantly consumed
Solution Approach 1:
The patent divides a single large MIP prediction matrix into multiple smaller sub-matrices. Each sub-matrix can be independently stored and processed, reducing the memory footprint while maintaining the overall prediction functionality. The segmentation allows the system to handle large prediction tasks through manageable matrix blocks without sacrificing prediction accuracy.
Solution Approach 2:
The patent transforms the traditional 2D matrix storage approach by introducing a hierarchical structure with multiple dimensions: sub-matrix indices, element indices within sub-matrices, and optional tile structures. This multi-dimensional organization enables efficient memory access patterns and reduces the contiguous memory requirement for storing large prediction matrices.
2Measurement precision
If original MIP prediction matrices are used, then prediction accuracy is maintained, but computational complexity increases
Solution Approach 1:
By segmenting the large prediction matrix into smaller sub-matrices, the patent enables parallel processing of multiple sub-matrices simultaneously. This segmentation reduces the computational burden on single processing units and allows for more efficient utilization of available computational resources, thereby reducing overall computational complexity while maintaining prediction accuracy.
Solution Approach 2:
The patent introduces dynamic sub-matrix selection based on block shape and prediction mode. Instead of always using fixed large matrices, the system dynamically selects appropriate sub-matrix sizes and configurations matched to the specific coding block characteristics, optimizing computational efficiency for each case while preserving prediction accuracy.
3Quantity of substance
If smaller matrices are used to reduce storage, then resource efficiency improves, but prediction accuracy deteriorates
Solution Approach 1:
The patent implements a nested structure where multiple levels of matrix decomposition are applied: large prediction matrices contain sub-matrices, which in turn contain individual elements that can be selectively processed. This nesting allows the system to use smaller matrices at appropriate granularities while combining their results to achieve the accuracy of larger matrices, effectively solving the storage-accuracy tradeoff.
Solution Approach 2:
The patent performs preliminary organization of prediction matrices into structured sub-matrices with specific patterns and relationships established in advance. This preliminary structuring enables the decoder to reconstruct accurate predictions by combining pre-organized sub-matrix results, maintaining prediction accuracy while using smaller, more efficiently stored matrix components.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
Video coding using reduced-rank matrices for matrix intra prediction (MIP). In a video decoding method, information is obtained indicating a matrix intra prediction mode for a current block. Based on the intra prediction mode and the current block shape, a first matrix and a second matrix are selected from a stored set of intra prediction matrices. Each of the matrices in the stored set of intra prediction matrices is associated with a respective block shape and intra prediction mode, and wherein all matrices associated with the same block shape have the same rank. The current block is predicted using a product of the first matrix, the second matrix, and a plurality of reference samples.