Affine Motion Compensation for Video Encoding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current video compression systems, such as HEVC and VVC, face inefficiencies in inter-prediction methods, particularly in handling complex motions like zoom, rotation, and irregular motions, due to limitations in motion representation and prediction accuracy.

Innovation Solution

The method involves using affine motion models and refined motion vector prediction techniques, including bilateral matching and sub-sample precision refinement, to improve motion compensation in inter-prediction, especially for blocks with specific shapes and sizes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If affine motion models with refined motion vector prediction are used, then motion representation accuracy for complex motions is improved, but encoding complexity and time overhead increase

Engineering Contradiction:
Improvemotion representation accuracyVSAvoidencoding complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the motion compensation process into distinct stages: initial motion vector prediction, bilateral matching refinement, and sub-sample precision refinement. Each stage processes specific candidate blocks independently, allowing parallel computation and reducing overall encoding complexity while maintaining high motion representation accuracy for complex motions like zoom and rotation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial refinement actions by selectively performing bilateral matching and sub-sample refinement only on candidate blocks that meet specific criteria (e.g., prediction error thresholds, block size, motion complexity). This partial application reduces the total computational load compared to refining all blocks, while still achieving sufficient motion representation accuracy for complex motions where needed.

Inventive Principle:
Principle #16Partial or excessive action

2Measurement precision

If bilateral matching and sub-sample precision refinement are performed for all candidate blocks, then prediction accuracy is improved, but encoding time increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidencoding speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent implements conditional refinement by evaluating prediction error metrics and block characteristics to determine whether bilateral matching and sub-sample refinement should be applied. Only candidate blocks exceeding certain error thresholds or exhibiting complex motion patterns undergo full refinement, while simpler blocks use faster prediction methods, thereby maintaining high prediction accuracy for difficult cases without proportionally increasing encoding time for all blocks.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent applies different levels of refinement quality to different candidate blocks based on their local characteristics. Blocks with complex motion patterns, larger sizes, or higher prediction errors receive full bilateral matching and sub-sample refinement, while blocks with simple motion or low error use simplified prediction. This local adaptation of refinement quality maintains overall prediction accuracy while significantly reducing total encoding time compared to uniform refinement of all blocks.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4398568A1Encoding/decoding video picture data
Publication Date: 2024.07.10 BEIJING XIAOMI MOBILE SOFTWARE CO LTD
  • EP4398568A1 patent drawingFigure 1~2
  • EP4398568A1 patent drawingFigure 3~4
  • EP4398568A1 patent drawingFigure 5

AI summary

The present application relates to a method of decoding a video picture by means of motion compensated temporal bi-prediction of an inter coded block using two reference video pictures in two separate reference video picture lists, and two affine motion fields defined by at least two control point motion vectors, denoted CPMVs, associated to each reference picture, refined CPMVs being obtained as output of the following first step (171) or, optionally of the second step (172) that follows the first step (171), - in said first step (171), for each CPMV, a bilateral matching being performed for a block centred on the CPMV to derive at least two refined CPMVs at an integer precision, and a set of CPMVs comprising either non-refined and/or refined CPMVs at an integer precision being selected to lead to an overall predicted block with a minimum bilateral matching cost, - in said second step (172), for each successive CPMV of the selected set of CPMVs that is associated with a block, each CPMV is refined at a sub-sample precision to minimize a bilateral matching cost of said block, wherein the second step (172) is bypassed according to a comparison between a bilateral matching cost associated with a CPMV that is performed in the first step and a threshold.