Neural Network Mesh Deformation Approximation for Real-Time Character Rigs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing computer animation technologies face challenges in achieving realistic mesh deformations for real-time applications due to computational constraints, as film-quality rigs are not optimized for interactive use and require significant computational resources.

Innovation Solution

A method that approximates mesh deformations using neural networks, splitting the deformation process into linear and nonlinear components, where linear deformations are computed directly from the skeleton transformations and nonlinear components are learned by neural networks, allowing for a computationally less expensive approximation that can be used in real-time interactive applications.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If film-quality character rigs are used for real-time applications, then mesh deformation realism is improved, but computational cost and evaluation time increase significantly

Engineering Contradiction:
Improvemesh deformation realismVSAvoidevaluation time
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The patent segments the mesh deformation computation into two distinct parts: linear deformation components that can be computed efficiently using traditional skeletal animation methods, and nonlinear deformation components that require more complex calculations. By separating these components, the system can apply different computational strategies to each, reducing the overall evaluation time while preserving visual quality.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent pre-computes and stores the nonlinear deformation characteristics during an offline training phase using neural networks. This preliminary action allows the system to learn complex deformation patterns in advance, so that during real-time execution, only the pre-trained neural networks need to be evaluated rather than computing full film-quality deformations from scratch.

Inventive Principle:
Principle #10Preliminary action

2Productivity

If computational constraints are imposed on character rigs for interactive applications, then real-time performance is improved, but mesh deformation realism deteriorates

Engineering Contradiction:
Improvereal-time performanceVSAvoidmesh deformation realism
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent changes the computational parameters by replacing the traditional computationally intensive deformation algorithm with a neural network-based approximation. The neural network is trained to replicate the behavior of film-quality rigs but with significantly reduced evaluation time, effectively changing the mathematical parameters from exact but slow computations to approximate but fast computations.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent creates a simplified copy of the film-quality rig behavior through neural networks. Instead of using the original complex deformation system, a trained neural network model is created that copies the essential visual characteristics of film-quality deformations while requiring far less computational resources during real-time execution.

Inventive Principle:
Principle #26Copying

3Productivity

If traditional linear deformation methods are used, then computational efficiency is improved, but deformation quality and realism deteriorate

Engineering Contradiction:
Improvecomputational efficiencyVSAvoiddeformation quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent creates a composite deformation system that combines linear deformation methods (which are computationally efficient) with neural network-based nonlinear deformation corrections (which add realism). This composite approach allows the system to benefit from both the speed of linear methods and the quality of complex nonlinear deformations.

Inventive Principle:
Principle #40Composite materials

Solution Approach 2:

The neural network acts as an intermediary between the simple linear deformation computation and the desired high-quality deformation result. The linear method provides a base deformation that is computationally efficient, and the neural network mediates by adding the necessary nonlinear corrections to achieve film-quality realism without paying the full computational cost of traditional methods.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3454302B1Approximating mesh deformation for character rigs
Publication Date: 2024.06.05 DREAMWORKS ANIMATION LLC
  • EP3454302B1 patent drawingFigure 1A
  • EP3454302B1 patent drawingFigure 1B
  • EP3454302B1 patent drawingFigure 2A~2B

AI summary

Systems, methods, and computer-readable medium for approximating mesh deformations for character rigs are disclosed. An embodiment includes applying a first deformation function to one or more mesh elements to determine an intermediate position based on a transform to a first structural element, wherein the one or more mesh elements are assigned to the first structural element, generating an offset based on a second deformation function for the one or more mesh elements using a deformation function approximation model, wherein the offset is a positional offset of the one or more mesh elements from the intermediate position to a target position corresponding to the transform applied to the first structural element, and generating a combined mesh deformation for the one or more mesh elements by combining the intermediate position and the offset.