Geometry Partition Coding With MMVD Motion Vector Signaling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video coding technologies face inefficiencies in predicting motion vectors, particularly in scenarios where block partitions require different motion vectors, leading to increased data requirements and reduced compression efficiency.

Innovation Solution

Implementing a geometry partition mode (GPM) and merge mode with motion vector difference (MMVD) to partition blocks into multiple parts, using angle and edge indexes for angles and edge positions, and distance and direction indexes for distances and directions, to determine motion vectors for each part, allowing for more precise motion vector prediction.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If motion vectors are predicted from surrounding blocks using conventional merge mode, then data requirements are reduced, but prediction accuracy deteriorates when block partitions require different motion vectors

Engineering Contradiction:
Improvedata required for motion vector signalingVSAvoidmotion vector prediction accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The current block is divided into two sub-blocks (first and second sub-blocks) with different motion vector predictions. Each sub-block can have its own motion vector difference (MVD) applied independently, allowing different motion characteristics to be captured for different regions while still using compact indexing to signal the parameters.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different motion vector prediction strategies are applied to different sub-blocks within the same current block. The first sub-block uses one prediction method while the second sub-block uses another, allowing each region to be optimized for its specific motion characteristics rather than applying a uniform approach.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If blocks are partitioned into multiple parts with different motion vectors, then motion prediction accuracy is improved, but data requirements increase

Engineering Contradiction:
Improvemotion vector prediction accuracyVSAvoiddata required for motion vector signaling
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

Multiple motion vector prediction candidates are merged into a single compact representation using the MMVD technique. Instead of signaling separate motion vectors for each sub-block, the invention uses a base motion vector plus a single motion vector difference (MVD) that is applied to both sub-blocks, significantly reducing the data required while maintaining prediction accuracy.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The invention changes the parameter representation from signaling complete motion vectors to signaling only the difference (delta) from a base motion vector. This parameter transformation reduces the number of bits required to encode motion information while preserving the ability to represent different motion characteristics across sub-blocks.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If conventional merge mode is used for inter prediction, then compression ratio is limited, but device complexity is reduced

Engineering Contradiction:
Improvecompression ratioVSAvoidprocessing complexity for motion vector prediction
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

Motion vector prediction candidates are prepared in advance and stored in a candidate list before the actual prediction process. The decoder can then select from these pre-computed candidates using simple indexing, reducing the computational complexity during real-time decoding while still achieving high compression ratios through efficient motion compensation.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4111684B1Geometry partition mode and merge mode with motion vector difference signaling
Publication Date: 2025.10.15 TENCENT AMERICA LLC
  • EP4111684B1 patent drawingFigure 1
  • EP4111684B1 patent drawingFigure 2
  • EP4111684B1 patent drawingFigure 3

AI summary

In a method of decoding, coded information of a block is received from a coded video bitstream. The coded information indicates that the block is partitioned based on a geometry partition mode (GPM) into a first part and a second part and inter predicted based on merge mode with motion vector difference (MMVD). That a motion vector for the first part of the block partitioned by GPM is not identical to a motion vector for the second part of the block partitioned by GPM is determined. The block is further reconstructed based on the motion vector for the first part of the block and the motion vector for the second part of the block.