Semantic-Map Image Generation for Rare Traffic Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining training images for image classifiers in various traffic situations is difficult and expensive, and manually labeling pixels in these images is labor-intensive, while synthetically generated data is underrepresented and requires improved realism for effective classification.
Innovation Solution
A method using a generator to create realistic images from semantic maps, combined with a discriminator to optimize image quality, and a mixed image training approach to enhance the training dataset for image classifiers, incorporating both actual and generated images to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If actual training images are collected from various traffic situations, then the classifier can learn from real data, but obtaining sufficient training images for rare traffic situations is difficult and expensive
Solution Approach 1:
The patent uses a generator to create synthetic training images that copy the essential characteristics of real traffic images. These generated images serve as substitutes for rare and expensive-to-obtain real training data, allowing the classifier to learn from both real and synthetic examples without requiring extensive manual data collection
Solution Approach 2:
The patent merges real training images with synthetically generated images to create an augmented training dataset. This combination allows the system to leverage the authenticity of real images while supplementing them with additional examples of rare traffic situations generated by the GAN, thereby improving classification accuracy without proportionally increasing data collection efforts
2Manufacturing precision
If training images are manually labeled with ground truth, then accurate training data is obtained, but the process requires lots of manual work
Solution Approach 1:
The patent copies semantic information from real training images to generate corresponding synthetic images with automatically transferred labels. The generator creates images that preserve the semantic structure and object relationships, allowing the ground truth labels to be applied to both real and generated images without additional manual labeling
Solution Approach 2:
The patent performs preliminary labeling on a subset of real images, then uses the trained generator to automatically generate additional training images with corresponding labels. This preliminary action reduces the overall manual labeling burden by creating a multiplier effect where one labeled image can generate multiple labeled training examples
3Ease of manufacture
If synthetically generated training data is used, then data collection is easier, but the generated images lack realism and require improved quality for effective classification
Solution Approach 1:
The patent implements a feedback mechanism using a discriminator that evaluates the realism of generated images and provides feedback to the generator. This adversarial process continuously refines the generator's output, improving image realism while maintaining the ease of synthetic data generation. The discriminator's feedback loop ensures that generated images become increasingly indistinguishable from real images
Solution Approach 2:
The patent employs a dynamic training process where the generator and discriminator are trained alternately in an adversarial manner. The generator dynamically improves its image generation capability in response to the discriminator's increasing ability to detect fakes, creating a moving target that continuously pushes the quality of generated images upward while maintaining computational efficiency
Data Source
AI summary
A method for training a generator for images from a semantic map that assigns each pixel of the image a semantic meaning of an object to which that pixel belongs. In the method, a mixed image is generated from an image generated by the generator and a determined actual training image, in which mixed image a first genuine subset of pixels is occupied by relevant corresponding pixel values of the image generated by the generator and the remaining genuine subset of pixels is occupied by relevant corresponding pixel values of the actual training image; and the images generated by the generator, the actual training image, and at least one mixed image, which belong to the same semantic training map, are supplied to a discriminator, which is configured to distinguish images generated by the generator from actual images of the scenery predefined by the semantic training map.