Composite Image Learning for Realistic Autonomous Driving Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The construction of training datasets for artificial intelligence models in autonomous driving requires significant time, cost, and manpower due to the need for large amounts of labeled real road images from various environments, and using simulation images results in poor model performance on real images due to feature discrepancies.
Innovation Solution
A learning device that generates composite images similar to real images by applying convolution operations and discrimination networks to simulation images, updating parameters to reflect real image features, and using a frequency domain for enhanced discrimination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If simulation road images are used to construct training datasets, then the cost and time of data construction are reduced, but the model performance on real road images deteriorates due to feature differences
Solution Approach 1:
The patent merges simulation road images and real road images to create composite training datasets. By combining the advantages of both simulation images (automatic labeling, cost efficiency) and real images (authentic features, visual characteristics), the system achieves both reduced construction time and improved model performance on real road images
Solution Approach 2:
The patent creates composite images that copy and integrate the essential visual features and characteristics of real road images while maintaining the structural benefits of simulation images. This allows the model to learn from authentic road scene features without incurring the high costs of manual real image labeling
2Manufacturing precision
If manual labeling is performed on real road images for multiple AI models, then the training dataset quality improves, but the cost and time consumption increase astronomically
Solution Approach 1:
The patent combines simulation images with automatic labeling capabilities and real image visual features to create composite training datasets. This merging approach maintains high labeling quality suitable for multiple AI models while dramatically improving construction efficiency by avoiding manual labeling of large numbers of real images
Solution Approach 2:
The system copies the essential visual characteristics and features of real road images into composite images generated from simulation data. This allows high-quality training data to be created for multiple AI models without the need for expensive manual labeling processes
Data Source
AI summary
A learning device for generating a composite image similar to a real image, a learning method, a testing device for generating a training dataset, and a testing method are provided. The learning device includes a composite image generation network that applies a convolution operation to a simulation image at least once to generate a composite image. The learning device has a first discrimination network that applies a first discrimination operation to a first input image to generate a first discrimination score and has a second discrimination network that applies a second discrimination operation to a second input image to generate a second discrimination score. The learning device also includes a parameter update device that performs back-propagation of a first loss generated with reference to at least a portion of the first discrimination score or the second discrimination score to update a parameter of the composite image generation network.