Synthetic Image Generator for Autonomous Driving Training Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional methods for training image classifiers for autonomous driving face challenges in obtaining representative training data, especially for rare traffic situations, and require significant manual effort for labeling, while existing synthetic data generation methods may not adequately address these issues.

Innovation Solution

A method utilizing a discriminator and generator, based on deep neural networks, to characterize semantic segmentation and style properties of input images, allowing for the generation of synthetic training data that can enhance the training of image classifiers, including the use of encoder-decoder devices and feedback mechanisms to improve training efficiency and stability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If real training images are collected from various traffic situations, then the training data represents real-world scenarios, but obtaining sufficient data for rare traffic situations is difficult and expensive

Engineering Contradiction:
Improverepresentativeness of training dataVSAvoidefficiency of training data acquisition
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent uses a generator network to create synthetic training images that copy the essential characteristics of real traffic scenes. The generator produces photorealistic images of rare traffic situations without requiring physical collection, thereby resolving the contradiction between data representativeness and acquisition efficiency

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system employs self-service through automated synthetic data generation. The generator and discriminator networks work together in an adversarial framework where the generator automatically creates diverse training samples without human intervention, eliminating the need for manual data collection and labeling efforts

Inventive Principle:
Principle #25Self-service

2Productivity

If synthetic training data is generated using GANs, then data acquisition efficiency improves, but the training process becomes complex and computationally demanding

Engineering Contradiction:
Improveefficiency of training data acquisitionVSAvoidcomplexity of training process
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent segments the complex training process into distinct functional components: an encoder that extracts features, a generator that creates images, and a discriminator that validates realism. This segmentation allows each component to be optimized independently and simplifies the overall training methodology

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The adversarial feedback mechanism between generator and discriminator provides automatic quality assessment during training. The discriminator's feedback signals guide the generator to improve image realism, creating a self-regulating training process that reduces the need for complex external validation procedures

Inventive Principle:
Principle #23Feedback

3Measurement precision

If manual labeling of training images is performed, then accurate ground truth is obtained, but significant manual work and time are required

Engineering Contradiction:
Improveaccuracy of ground truth labelingVSAvoidtime for manual labeling
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system copies semantic information from simplified source data into photorealistic generated images. The encoder processes simple semantic maps and the generator reproduces realistic images with corresponding ground truth automatically embedded, eliminating manual labeling while preserving accuracy

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces the mechanical process of manual labeling with an automated neural network pipeline. The encoder-generator system automatically generates both images and their corresponding ground truth labels through computational processes, substituting human manual work with algorithmic automation

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

Data Source

PatentEP4495890A1Method and apparatus for processing image data
Publication Date: 2025.01.22 ROBERT BOSCH GMBH
  • EP4495890A1 patent drawingFigure 1~2
  • EP4495890A1 patent drawingFigure 3~5
  • EP4495890A1 patent drawingFigure 6

AI summary

Method, for example a computer-implemented method, for processing image data characterizing, for example at least, an input image, comprising: providing the input image, determining first information that characterizes a semantic segmentation of the input image by means of a discriminator, determining second information that is different from the first information and that characterizes properties associated with the input image by means of the discriminator.