Synthetic Training Data Generation from 3D Models and Auto-Annotation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for generating training data for machine learning computer vision models are time-consuming and require significant expertise, involving manual data collection and labeling, which limits accessibility and efficiency.

Innovation Solution

A computer-implemented method and system for automatically generating synthetic training data sets using user-defined 2D or 3D models, allowing users to input parameters for rendering and annotation, thereby creating photorealistic images with minimal human effort and expertise.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual data collection and labeling is used, then data quality and accuracy are improved, but time consumption and effort increase significantly

Engineering Contradiction:
Improvedata qualityVSAvoidtime consumption
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent uses synthetic data generation by copying and rendering 3D model properties (geometry, materials, lighting) to create realistic 2D images with automatic annotations, eliminating manual data collection while maintaining data quality through physics-based rendering engines

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces manual mechanical labeling processes with automated computer vision algorithms that extract annotations directly from 3D model data and rendering parameters, substituting human effort with computational processes

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If manual labeling and data preparation is performed, then training data accuracy is improved, but complexity of the process increases

Engineering Contradiction:
Improvetraining data accuracyVSAvoidprocess complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges data generation, annotation creation, and training preparation into a single integrated synthetic data pipeline, where 3D model rendering automatically produces both images and corresponding annotations without separate manual steps

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The system performs self-service by automatically generating annotated training data from 3D models without requiring external human annotators, using the rendering engine and scene configuration to directly produce labeled images

Inventive Principle:
Principle #25Self-service

3Reliability

If specialized knowledge in computer vision is required, then model training effectiveness is improved, but ease of operation decreases

Engineering Contradiction:
Improvemodel training effectivenessVSAvoidease of use
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The patent introduces an intermediary synthetic data generation system that translates user-friendly 3D model inputs into professionally-quality training data, mediating between simple user operations and complex training requirements without requiring users to understand computer vision algorithms

Inventive Principle:
Principle #24Intermediary (Mediator)

4Quantity of substance

If extensive manual effort is spent on data collection, then quantity of training data is improved, but productivity decreases

Engineering Contradiction:
Improvequantity of training dataVSAvoidproductivity
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The patent uses periodic rendering of 3D models with varied parameters (lighting conditions, camera angles, object positions) to systematically generate large quantities of diverse training data through automated batch processing, increasing productivity through systematic repetition

Inventive Principle:
Principle #19Periodic action

Solution Approach 2:

The system performs preliminary action by pre-defining 3D models, scene configurations, and rendering parameters before data generation, allowing bulk production of training data without repeated manual setup for each image

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4105826A1Computer-implemented method and system for generating a synthetic training data set for training a machine learning computer vision model
Publication Date: 2022.12.21 VOLKSWAGEN AG
  • EP4105826A1 patent drawingFigure 1
  • EP4105826A1 patent drawingFigure 2
  • EP4105826A1 patent drawingFigure 3

AI summary

Computer-implemented method for generating a synthetic training data set for training a machine learning computer vision model for performing at least one user defined computer vision task, in which spatially resolved sensor data are processed and evaluated with respect to at least one user defined object of interest, including: - receiving at least one, in particular 2D or 3D, model (10) of a user defined object of interest based on user input data; - determining at least one render parameter (60) and preferably a plurality of render parameters (56, 58, 60, 62, 64, 66) based on user input data; - generating a set of training images (12, 13) by rendering the at least one model (10) of the object of interest based on the at least one render parameter (60); - generating annotation data for the set of training images (11) with respect to the at least one object of interest; - providing a training data set comprising the set of training images (12, 13) and the annotation data for being output to the user and/or for training the computer vision model.