Affine Motion Vector Refinement With Reused Subblock Motion Fields
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Certain decoder side motion vector refinement techniques for bi-directional predicted affine merge candidates add complexity and consume excessive processing power in video decoders.
Innovation Solution
The proposed techniques generate subblock motion fields only once for each block, reducing the number of times they are generated across possible offsets, thereby simplifying the process and conserving processing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If decoder side motion vector refinement is performed for bi-directional predicted affine merge candidates using conventional techniques, then motion vector accuracy is improved, but device complexity and processing power consumption increase significantly
Solution Approach 1:
The current block is divided into multiple subblocks, and motion vector refinement is performed independently for each subblock. This segmentation allows the complex refinement process to be broken down into smaller, more manageable units that can be processed in parallel, reducing overall decoder complexity while maintaining accuracy.
Solution Approach 2:
Initial motion vectors are generated for all subblocks before the refinement process begins. These preliminary motion vectors serve as starting points for the bilateral matching process, allowing the refinement algorithm to focus only on optimizing from these pre-computed values rather than searching from scratch, thereby reducing processing requirements.
2Measurement precision
If subblock motion fields are generated multiple times for each offset during DMVR, then motion vector refinement accuracy is improved, but processing power and time consumption increase
Solution Approach 1:
The motion field for the first prediction direction is generated once in advance before the bilateral matching process begins. This pre-generated motion field is then reused across multiple offset evaluations, eliminating the need to regenerate it repeatedly and significantly improving processing efficiency.
Solution Approach 2:
The initially generated motion field serves multiple purposes: it is used as the basis for bilateral matching across all offsets and also serves as one of the two fields needed for computing the refined motion vector. This multi-functional use of a single generated field reduces redundant computations.
3Measurement precision
If conventional DMVR techniques are applied to bi-directional predicted affine merge candidates, then prediction accuracy is improved, but power consumption increases excessively
Solution Approach 1:
Processing is segmented into independent subblock operations that can be executed efficiently with minimal data movement and memory access. This segmentation reduces the overall computational burden and power consumption by allowing localized processing with smaller working sets.
Solution Approach 2:
Motion fields and initial computations are performed once in advance, and their results are cached and reused throughout the refinement process. This eliminates redundant computations that would otherwise consume excessive power, while still achieving accurate refined motion vectors through the bilateral matching process.
Data Source
AI summary
An example device includes one or more processors configured determine a plurality of subblocks for a current block of video data. For each subblock, the one or more processors (a) generate initial motion vectors for a first prediction direction and a second prediction direction according to an affine motion model, and (b) determine, based on the initial motion vectors, a subblock bilateral matching cost for each respective offset among a plurality of offsets. For each respective offset, the one or more processors determine a respective summation of subblock bilateral matching costs. The one or more processors determine a lowest summation of subblock bilateral matching costs. The one or more processors select an offset associated with the lowest summation of subblock bilateral matching costs. The one or more processors modify the affine motion model based on the selected offset and code the current block based on the modified affine motion model.