Video Encoding Linear Model Parameter Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression systems, such as HEVC and under-development VVC, face inefficiencies in determining spatial neighboring templates for linear model parameters, leading to sub-optimal prediction accuracy, especially in cases of rectangular CUs with unbalanced dimensions, non-rectangular predictions, and multi-hypothesis predictions.
Innovation Solution
A method for video encoding and decoding that determines a spatial neighboring template with a number of reconstructed samples corresponding to 2k, where k is the maximum integer smaller than the sum of the block width and height, and adjusts the relative position and sub-sampling of samples to optimize the selection of reference samples for linear model parameters, including adaptive selection based on block geometry and coding modes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of reconstructed samples in the spatial neighboring template is increased to improve prediction accuracy, then the linear model parameter accuracy improves, but the computational complexity increases
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the number of reconstructed samples (n=2^k) based on block geometry characteristics. The parameter k is selected according to the sum of block width and height, allowing the system to adapt the sample count to match the specific block size, thereby optimizing the balance between prediction accuracy and computational load for different block types.
Solution Approach 2:
The patent implements dynamics by making the template sample selection adaptive rather than fixed. The number of samples n=2^k is dynamically determined based on the block dimensions (width + height), enabling the system to automatically adjust the complexity of the linear model parameter calculation to suit the specific characteristics of each block being encoded or decoded.
2Device complexity
If a fixed number of samples is used in the spatial neighboring template, then the computational complexity is reduced, but the prediction accuracy deteriorates for blocks with unbalanced dimensions
Solution Approach 1:
The patent applies local quality by tailoring the number of template samples to the specific characteristics of each block. Instead of using a uniform sample count for all blocks, the method selects n=2^k samples based on the local block geometry (width + height), ensuring that each block receives an appropriate number of samples suited to its dimensions, thereby improving prediction accuracy for unbalanced blocks without uniformly increasing complexity.
3Measurement precision
If the spatial neighboring template is optimized for rectangular blocks, then the prediction accuracy improves for rectangular CUs, but the adaptability to non-rectangular and multi-hypothesis predictions deteriorates
Solution Approach 1:
The patent applies universality by designing a sample selection method (n=2^k based on width+height) that works across multiple prediction types including rectangular CUs, non-rectangular blocks, and multi-hypothesis predictions. This universal approach allows the same template optimization strategy to benefit various block types and prediction modes, enhancing both accuracy and adaptability simultaneously.
Data Source
AI summary
Different implementations are described, particularly implementations for video encoding and decoding based on a linear model responsive to neighboring samples are presented. Accordingly, for a block being encoded or decoded in a picture, at least one spatial neighboring template is determined and at least one linear model parameter is determined based on reconstructed samples of the at least one spatial neighboring template. In a first embodiment, the number N of reconstructed samples corresponds to N=2k with k chosen so that n is the maxim integer smaller than sum of the block width and block height. In a second embodiment, the an offset for the relative position of a first sample in the template among samples of a left, respectively top, neighboring line of the block is determined. In a third embodiment, the number of reconstructed samples is set to a higher number in the larger dimension of the block.