Motion Vector Sign Prediction Using Template Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In High-Efficiency Video Coding (HEVC), the existing motion vector prediction methods for Inter Advanced Motion Vector Prediction (AMVP) mode are inefficient in coding motion vector signs, leading to increased bitstream size and computational complexity due to separate coding of magnitude and sign components.
Innovation Solution
A method for motion vector reconstruction using template matching between neighboring reconstructed pixels and candidate motion vector positions, where the correct sign pair is determined based on performance criteria such as SAD or SSD, and optionally using decoder-side statistics to improve accuracy and reduce computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If separate coding of magnitude and sign components is used in existing motion vector prediction methods, then motion vector coding can be performed with standard AMVP mode, but bitstream size increases and computational complexity increases
Solution Approach 1:
The patent combines the magnitude and sign components of motion vector differences into a unified coding scheme. Instead of separately encoding magnitude and sign, the invention uses a single motion vector difference value that inherently contains both magnitude and direction information, thereby reducing bitstream size while maintaining coding efficiency.
Solution Approach 2:
The invention changes the parameter representation from separate magnitude and sign components to a unified motion vector difference parameter. This parameter transformation allows the motion vector to be encoded more efficiently by exploiting the statistical properties of motion vector signs through template matching, reducing the overall bitstream requirement.
2Ease of operation
If separate coding of magnitude and sign components is used in existing motion vector prediction methods, then motion vector coding can be performed with standard AMVP mode, but computational complexity increases
Solution Approach 1:
The patent combines the magnitude and sign components of motion vector differences into a unified coding scheme. Instead of separately encoding magnitude and sign, the invention uses a single motion vector difference value that inherently contains both magnitude and direction information, thereby reducing bitstream size while maintaining coding efficiency.
Solution Approach 2:
The invention employs template matching that uses already-decoded neighboring pixel values to predict motion vector signs. This self-service approach leverages existing decoded data to improve coding efficiency without requiring additional complex processing or external information, thereby reducing computational complexity.
3Measurement precision
If template matching is performed for all candidate reconstructed MV positions, then decoding accuracy improves, but computational complexity increases
Solution Approach 1:
The patent applies template matching selectively based on local characteristics of the video content and motion patterns. By adapting the template matching process to local regions and using already-decoded neighboring pixels, the invention achieves high decoding accuracy in areas where it matters most while avoiding unnecessary computations in other regions.
Solution Approach 2:
The invention performs template matching for a limited set of candidate positions rather than all possible positions. By focusing computational effort on the most likely candidate reconstructed MV positions based on MVP and MVD magnitude, the patent achieves sufficient decoding accuracy without the excessive computational cost of evaluating all possible positions.
Data Source
AI summary
A method and apparatus of motion vector reconstruction based on one MVP (motion vector predictor) and one MVD (motion vector difference) are disclosed. According to this method, input data associated with a video sequence including a current block in a current picture are received. Candidate reconstructed MV (motion vector) positions corresponding to all candidate sign pairs of the MVD are derived, where each candidate reconstructed MV position is derived as a vector sum of the MVP and the first magnitude and the second magnitude of the MVD with one candidate sign pair of the MVD. Template matching between the template for the current block and a corresponding template for each candidate reconstructed MV position is evaluated. A target candidate MV position among all candidate reconstructed MV positions is selected. The current block is encoded or decoded using a target candidate MV corresponding to the target candidate MV position.