Affine Motion Model for Frame Rate Up-Conversion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current frame rate up-conversion techniques, such as bilateral matching, increase computation complexity without reducing bitstream size, and fail to effectively handle non-linear object motions like rotation and zooming, which are common in video content.

Innovation Solution

The implementation of an affine motion model in frame rate up-conversion, which uses a simplified motion vector field determined by control point vectors to capture non-linear motions, reducing computational complexity and bitstream size while improving coding efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If bilateral matching is used for frame rate up-conversion, then video quality is improved, but computation complexity increases without bitstream size reduction

Engineering Contradiction:
Improvevideo qualityVSAvoidcomputation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the motion representation into control point motion vectors (CPMVs) that define an affine transformation field. Instead of computing bilateral matching for all pixels, the method divides the block into regions governed by control points, reducing computation while maintaining quality through affine motion modeling.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the motion model parameters from simple translational vectors to affine transformation parameters (CPMVs). This parameter change enables the model to represent complex non-linear motions (rotation, zooming) with fewer parameters, reducing computation complexity while improving video quality through more accurate motion compensation.

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If translational motion model is used, then computation is simpler, but non-linear motions like rotation and zooming are not effectively handled

Engineering Contradiction:
Improvecomputation simplicityVSAvoidmotion model accuracy
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent introduces dynamic affine transformation parameters (CPMVs) that can adapt to different motion patterns within a block. The control point motion vectors enable the motion model to dynamically adjust to non-linear motions such as rotation and zooming, making the system versatile while maintaining computational efficiency through parameterized transformation.

Inventive Principle:
Principle #15Dynamics

3Productivity

If affine motion model is applied to capture non-linear motions, then coding efficiency is improved, but bitstream size may increase

Engineering Contradiction:
Improvecoding efficiencyVSAvoidbitstream size
Core Design Contradiction:
ProductivityVSQuantity of substance

Solution Approach 1:

The patent applies partial affine transformation by using a limited number of control points (typically 2-4 per block) rather than full affine modeling for the entire block. This partial application captures the essential non-linear motion characteristics while keeping the bitstream overhead minimal, as only the CPMVs for control points need to be encoded rather than parameters for all pixels.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3520406B1Improvements on frame rate up-conversion coding mode
Publication Date: 2023.02.01 QUALCOMM INC
  • EP3520406B1 patent drawingFigure 1
  • EP3520406B1 patent drawingFigure 2A~2B
  • EP3520406B1 patent drawingFigure 3A~3B

AI summary

Methods, apparatuses, and computer-readable medium are provided for a frame rate up-conversion coding mode, in which an affine motion model is applied when conducting bilateral matching. The frame rate up-conversion coding mode can include generated additional frames from frames provided in a bitstream. In various implementations, bilateral matching includes, for a current block in a frame that is being generated, identifying a first block in a first reference picture a second block in a second reference picture. Affine (e.g., non-linear) motion information can be determined as between the first block and the second block. The current block can be predicted using the affine motion information.