Affine Motion Prediction Using Feature Points

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video coding technologies face challenges in achieving high-precision prediction, particularly when using affine motion models, due to the low precision of predicting current blocks based on reference blocks.

Innovation Solution

The proposed method improves prediction precision by calculating motion information for each subblock in the current block based on the location and motion information of at least two feature points in the reference block, ensuring a high degree of matching between the used location and motion information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional affine motion model is used for predicting current block from reference block, then prediction can be performed, but prediction precision is relatively low

Engineering Contradiction:
Improveprediction precisionVSAvoidmotion information calculation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The current block is divided into multiple subblocks, and motion information is calculated separately for each subblock based on corresponding subblocks in the reference block. This segmentation allows for more precise local motion compensation while maintaining manageable computational complexity through systematic processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different motion information is calculated for different subblocks within the current block, allowing each region to have optimized motion parameters tailored to its specific characteristics. This local optimization improves overall prediction precision by accounting for spatial variations in motion patterns.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If block-based motion compensation based on translational motion model is used, then computation is simplified, but it cannot satisfy high-precision prediction requirement in some scenarios

Engineering Contradiction:
Improveprediction precisionVSAvoidmotion model adaptability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The motion model transitions from a static translational model to a dynamic affine model that can adapt to different motion patterns. The affine transformation parameters are calculated based on control points, allowing the model to dynamically adjust to various motion scenarios including rotation, scaling, and shearing, thereby improving both precision and adaptability.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3557874B1Image prediction method and related device
Publication Date: 2025.04.23 HUAWEI TECH CO LTD
  • EP3557874B1 patent drawingFigure 1-a~1-b
  • EP3557874B1 patent drawingFigure 1-c~2
  • EP3557874B1 patent drawingFigure 3-a~3-b

AI summary

A picture prediction method and a related apparatus are disclosed. The picture prediction method includes: determining a reference block of a current block, where the reference block and the current block are spatially adjacent, and the reference block is predicted by using an affine motion model; obtaining location information and motion information of at least two feature points in the reference block, where the at least two feature points are located in subblocks in which at least two control points in the reference block are located, the at least two control points are control points used in the affine motion model used to predict the reference block, and motion information of the subblock in which the feature point is located is obtained based on the motion information of the corresponding feature point; calculating motion information of each subblock in the current block based on the location information and the motion information of the at least two feature points in the reference block; and performing motion compensation prediction on each subblock in the current block based on the motion information of each subblock in the current block, to obtain a prediction block of the current block. The solutions in the embodiments of the present invention help improve current block prediction precision.