Georeferenced Environment Rendering for Automated Training Image Labeling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for generating training image sets for machine vision systems, such as those used in UAVs and self-driving vehicles, are time-consuming and rely heavily on manual semantic segmentation by operators, making them impractical for large datasets.

Innovation Solution

A method and system that automatically generate training image sets by creating 2-D rendered images from georeferenced models of environments, associating them with labels, and linking them to native images, using algorithms to precisely segment environmental features without human intervention.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual semantic segmentation is used to generate training images, then labeling accuracy can be maintained, but the process becomes extremely time-consuming and impractical for large datasets

Engineering Contradiction:
Improvelabeling accuracyVSAvoidtime consumption
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent creates synthetic training images by rendering 2-D representations from 3-D environmental models. These synthetic images serve as copies that replicate real-world scenarios without requiring manual segmentation of actual photographs, thereby maintaining labeling accuracy while eliminating time-consuming manual processes.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system automatically generates labeled training images through algorithmic processing of environmental models. The computer system performs semantic segmentation and label generation autonomously without human intervention, making the process self-service and highly efficient for large datasets.

Inventive Principle:
Principle #25Self-service

2Adaptability or versatility

If manual semantic segmentation is performed by operators, then flexible handling of complex features is possible, but the process relies on operator skill and becomes impractical for thousands of images

Engineering Contradiction:
Improvehandling flexibilityVSAvoidprocessing throughput
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

Synthetic training images are generated by rendering from 3-D environmental models, creating accurate representations of complex environmental features without requiring manual interpretation. This copying approach maintains adaptability to complex scenarios while enabling automated processing of large volumes of images.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces the mechanical process of manual segmentation with automated computer-based rendering and image generation. The system uses algorithms to automatically create labeled training images, substituting human operator skills with computational processes that can handle thousands of images efficiently.

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

3Quantity of substance

If large datasets of thousands of training images are generated manually, then comprehensive training coverage is achieved, but the manual process becomes impractical and unsustainable

Engineering Contradiction:
Improvedataset sizeVSAvoidgeneration feasibility
Core Design Contradiction:
Quantity of substanceVSEase of manufacture

Solution Approach 1:

The system generates synthetic training images by copying and rendering environmental models from multiple viewpoints and conditions. This approach enables the creation of large datasets with comprehensive training coverage while maintaining ease of generation through automated computational processes.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The environmental models serve as universal sources from which numerous training images can be generated through automated rendering. A single 3-D model can produce multiple 2-D training images with different viewpoints, lighting conditions, and environmental variations, enabling comprehensive dataset creation without proportional increase in manual effort.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3792822B1Systems and methods for automatically generating training image sets for an environment
Publication Date: 2025.12.31 THE BOEING CO
  • EP3792822B1 patent drawingFigure 1~2B
  • EP3792822B1 patent drawingFigure 3A
  • EP3792822B1 patent drawingFigure 3B

AI summary

A computer-implemented method for generating a training set of images and labels for a native environment includes receiving physical coordinate sets, retrieving environmental model data corresponding to a georeferenced model of the environment, and creating a plurality of two-dimensional (2-D) rendered images each corresponding to a view from one of the physical coordinate sets. The 2-D rendered images include one or more of the environmental features. The method also includes generating linking data associating each of the 2-D rendered images with (i) labels for the one or more included environmental features and (ii) a corresponding native image. Additionally, the method includes storing the training set including the 2-D rendered images, labels, corresponding native images, and linking data.