Video Coding Merge Candidate List Generation Using Spatial Search Rounds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video coding methods, such as HEVC and VVC, face limitations in generating diverse spatial motion vector predictors for inter-picture prediction, which can lead to suboptimal performance in merge mode and motion vector prediction, especially when dealing with extended spatial merge candidates.
Innovation Solution
The proposed method involves performing multiple search rounds to identify motion vector predictors for adjacent and non-adjacent spatial candidate blocks, generating a merge candidate list that includes predictors from these searches, allowing for a more varied search pattern and improved prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple search rounds are performed to identify motion vector predictors from adjacent and non-adjacent spatial candidate blocks, then the diversity of search patterns for spatial merge candidates is enhanced and prediction accuracy is improved, but the computational complexity and encoding/decoding time increase
Solution Approach 1:
The search process is divided into multiple search rounds, where each round identifies motion vector predictors from specific spatial candidate blocks. The first search round identifies predictors from adjacent spatial candidate blocks (A, B, C, D, and E), while subsequent search rounds identify predictors from non-adjacent spatial candidate blocks (Ai, Bi, Ci, Di, and Ei). This segmentation allows systematic exploration of different spatial regions without overwhelming computational complexity in a single step.
Solution Approach 2:
The first search round performs preliminary identification of motion vector predictors from adjacent spatial candidate blocks before conducting subsequent search rounds for non-adjacent blocks. This preliminary action establishes a foundation of reliable predictors from nearby blocks, which can be combined with predictors from farther blocks to create a comprehensive merge candidate list with improved diversity and accuracy.
2Reliability
If up to seven merge candidates are used including ATMVP and STMVP modes, then the prediction performance is improved, but the number of RD checks and processing overhead increase
Solution Approach 1:
The merge candidate list is constructed dynamically by performing multiple search rounds and selectively including motion vector predictors from adjacent and non-adjacent spatial candidate blocks. The number of merge candidates is adapted based on the availability and quality of predictors from different spatial regions, allowing the system to use up to seven candidates (including ATMVP and STMVP modes) when beneficial, while potentially using fewer candidates when appropriate, thus balancing prediction performance with processing efficiency.
3Measurement precision
If all bins of merge index are context coded by CABAC, then the encoding precision is improved, but the encoding complexity and time consumption increase compared to context by-pass coding
Solution Approach 1:
The coding method for merge index bins is changed from uniform context by-pass coding to context-adaptive binary arithmetic coding (CABAC) for all bins. This parameter change in the coding strategy improves encoding precision by adapting the coding context to the specific characteristics of each merge index bin, allowing more efficient representation of the merge candidate selection information.
Data Source
AI summary
An apparatus and method for generating a merge candidate list for encoding or decoding of a video sequence includes performing a first search round. A first set of motion vector predictors associated with a set of adjacent spatial candidate blocks (A, B, C, D, and E) is identified based on the first search round. An n-th search round is performed. An n-th set of motion vector predictors associated with a set of non-adjacent spatial candidate blocks (Ai, Bi, Ci, Di, and Ei) is identified based on the n-th search round. The merge candidate list is generated using the first set of motion vector predictors and the n-th set of motion vector predictors.