MMVD Candidate Refinement for Accurate Motion Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding technologies face inefficiencies in predicting motion vectors, particularly in intra and inter-picture prediction, leading to suboptimal compression ratios and increased data requirements.
Innovation Solution
The implementation of motion vector difference (MMVD) candidate refinement methods, which involve generating refined motion vectors using fractional offsets and multiple reference pictures, to improve prediction accuracy and reduce redundancy in video data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional motion vector prediction methods are used, then device complexity is reduced, but prediction accuracy deteriorates leading to suboptimal compression ratios
Solution Approach 1:
The motion vector prediction process is segmented into multiple stages: generating multiple MMVD candidates from different reference pictures, calculating template matching costs for each candidate, and selectively refining only the most promising candidates. This segmentation allows the system to achieve high prediction accuracy while controlling computational complexity through hierarchical processing.
Solution Approach 2:
Instead of refining all motion vector candidates equally, the system applies partial refinement only to candidates that meet specific criteria (e.g., lowest template matching costs). This partial action approach concentrates computational resources on the most promising candidates, improving prediction accuracy without proportionally increasing overall device complexity.
2Measurement precision
If multiple reference pictures are used for MMVD candidate generation, then prediction accuracy is improved, but data processing requirements increase
Solution Approach 1:
The system performs preliminary template matching cost calculations on MMVD candidates before applying refinement. By pre-evaluating candidates using template matching, the system identifies and prioritizes the most promising candidates for refinement, reducing the amount of data that requires intensive processing while maintaining high prediction accuracy.
Solution Approach 2:
Different reference pictures are used for different MMVD candidates, with each candidate associated with specific reference picture(s) based on local characteristics. This local quality approach allows the system to optimize prediction for each candidate using the most appropriate reference data, improving accuracy without uniformly increasing data processing requirements across all candidates.
3Productivity
If MMVD candidate refinement is applied, then compression ratio is optimized, but computational complexity increases
Solution Approach 1:
The system uses template matching costs as feedback to guide the refinement process. Candidates with lower template matching costs are selected for refinement, and the refinement process itself generates feedback that further optimizes the selection. This feedback mechanism ensures that computational resources are allocated to refinement operations that will most improve compression ratio, optimizing the trade-off between complexity and productivity.
Solution Approach 2:
The refinement process is made dynamic by adaptively selecting which candidates to refine based on template matching results and other criteria. The system dynamically adjusts the refinement strategy, applying refinement only when and where it will most benefit compression performance, rather than uniformly applying refinement to all candidates. This dynamic approach optimizes compression ratio while controlling computational complexity.
Data Source
AI summary
Aspects of the disclosure provide methods and apparatuses for video encoding/decoding. In some examples, an apparatus for video decoding includes processing circuitry. The processing circuitry extracts, from a bitstream, merge with motion vector difference (MMVD) candidate information for a current block in a current picture. The processing circuitry generates a first MV refinement offset associated with a first motion vector for the MMVD candidate based on a refined step size and a plurality of refinement positions. The processing circuitry derives a first refined motion vector (MV) value associated with a MMVD candidate according to the MMVD candidate information and the generated first MV refinement offset. The processing circuitry reconstructing the current block according to a first reference block in a first reference picture, the first reference block is indicated by the derived first refined MV value.