Generator Rating for Realistic Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training image classifiers lack effective generation of realistic images, particularly for underrepresented scenarios like nighttime construction sites, leading to inadequate training data for automated driving systems.
Innovation Solution
A method for quantitatively rating generators that produce target segmentation maps, allowing for the creation of realistic images by combining a well-rated generator for segmentation maps with another generator to produce images that match the desired content and variability, using a predetermined class distribution to ensure coverage of rare scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a generator is used to synthetically generate training images for underrepresented scenarios, then the quantity of training data is increased, but the realism and quality of generated images deteriorate
Solution Approach 1:
The patent implements a feedback mechanism where a discriminator network evaluates the realism of generated images and provides feedback signals to the generator. This adversarial training process allows the generator to iteratively improve image quality while maintaining synthetic data generation, resolving the contradiction between quantity and realism.
Solution Approach 2:
The discriminator acts as an intermediary between the generator and the training process. It mediates the quality assessment of generated images and translates this into training signals for the generator, enabling the system to produce realistic images without requiring manual quality assessment.
2Adaptability or versatility
If generators are trained to produce diverse images covering rare scenarios, then the adaptability of training data is improved, but the measurement precision of image quality deteriorates
Solution Approach 1:
The patent employs dynamic class distribution sampling that adapts during training to prioritize underrepresented scenarios. The sampling strategy evolves from uniform to focused on rare classes, allowing the system to improve coverage of rare scenarios while maintaining quality assessment through the discriminator's learned evaluation metrics.
Data Source
AI summary
A method for quantitatively rating a trained generator that generates, from an input vector from a predetermined input distribution in connection with a target segmentation map, an image whose semantic content is in line with this target segmentation map. The method includes: drawing at least one input vector from the predetermined input distribution; drawing at least one list of classes from a predetermined class distribution; determining a target segmentation map, which assigns classes from this list of classes to the pixels of the image to be generated; generating, by means of the generator, an image from the input vector and the target segmentation map; determining a semantic segmentation map of the image by means of an image classifier; determining, using a predetermined metric, the degree of matching between this semantic segmentation map and the target segmentation map; determining, using this degree of matching, the quantitative rating of the generator.