DMVR Inter Prediction with Selective Motion Vector Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing demand for high-resolution and high-quality images and videos, along with the rise of immersive media formats, necessitates a high-efficiency image and video compression technology to reduce transmission and storage costs while maintaining quality.
Innovation Solution
Implementing a decoder-side motion vector refinement (DMVR) and bi-directional optical flow (BDOF) for inter prediction, with conditions to determine when these methods are applied, to enhance coding efficiency and reduce complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If DMVR and BDOF are applied to refine motion vectors, then prediction accuracy and compression efficiency are improved, but calculation complexity increases
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the refinement search range and step size based on motion vector magnitude and block size. When motion vectors are small or blocks are large, the search range is reduced and step size is increased, decreasing computational complexity while maintaining prediction accuracy for cases where it matters most.
Solution Approach 2:
The patent implements partial action by applying DMVR and BDOF only to specific blocks based on motion vector thresholds and block size criteria, rather than uniformly to all blocks. This selective application reduces overall computational complexity while maintaining high compression efficiency for the most beneficial cases.
2Measurement precision
If motion vector refinement is applied to all blocks, then prediction accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies motion vector refinement partially based on motion vector magnitude thresholds and block size criteria. Blocks with small motion vectors or large dimensions are selected for refinement, as these are the cases where prediction accuracy improvement is most significant, while skipping refinement for blocks where it would be less beneficial and time-consuming.
Solution Approach 2:
The patent dynamically changes refinement parameters including search range and step size based on motion vector magnitude and block size. This adaptive approach ensures that processing time is allocated efficiently to blocks that benefit most from refinement, reducing overall processing time while maintaining high prediction accuracy for critical cases.
3Adaptability or versatility
If combined inter and intra prediction mode is used, then prediction flexibility is improved, but DMVR and BDOF cannot be applied
Solution Approach 1:
The patent applies different prediction strategies to different blocks based on their characteristics. Blocks using combined inter and intra prediction modes receive standard prediction processing, while blocks using pure inter prediction modes receive DMVR and BDOF refinement. This local differentiation optimizes coding efficiency for each block type without compromising overall prediction flexibility.
Solution Approach 2:
The patent dynamically selects the appropriate prediction mode and refinement strategy based on block characteristics and motion vector properties. The system adapts its behavior in real-time, applying DMVR and BDOF only when beneficial, thereby maintaining prediction flexibility while maximizing coding efficiency for suitable cases.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image decoding method according to the present document comprises the steps of: determining whether or not an application condition of decoder-side motion vector refinement (DMVR) for applying motion vector refinement for a current block is satisfied; deriving a minimum sum of absolute differences (SAD) on the basis of L0 and L1 motion vectors of the current block if the application condition of the DMVR is satisfied; deriving refined L0 and L1 motion vectors of the current block on the basis of the minimum SAD; deriving prediction samples of the current block on the basis of the refined L0 and L1 motion vectors; and generating reconstructed samples of the current block on the basis of the prediction samples. With respect to whether or not the application condition of the DMVR is satisfied, the application condition of the DMVR is determined to be satisfied if a prediction mode, in which inter-prediction and intra-prediction are combined, is not applied to the current block.