Synthetic Training Data Generation for ML Classifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning classifiers require a large amount of labeled data for training, which is time-consuming and expensive due to the need for manual labeling of sensor data from various perspectives, lighting conditions, and locations.
Innovation Solution
A system that generates training data by modifying a base scene with changes in lighting, adding or replacing elements, and placing virtual sensors to create derivative scenes, allowing for automated generation of labeled training data without additional manual labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling of sensor data is performed to train machine learning classifiers, then the quality and accuracy of training data is improved, but the time and cost required for data collection and labeling increases significantly
Solution Approach 1:
The patent creates synthetic copies of sensor data by generating derivative scenes from a single labeled scene. Virtual sensors capture images from multiple positions, orientations, and lighting conditions, producing training data without requiring physical collection or manual labeling of each sample. This copying approach maintains data quality while dramatically reducing time and resource requirements.
Solution Approach 2:
The patent performs preliminary labeling of a single base scene, then uses automated processing to generate all derivative training samples from this pre-labeled source. The labeling action is performed once in advance rather than repeatedly for each training sample, significantly reducing the time and cost of data preparation while ensuring consistent labeling quality across all derived images.
2Adaptability or versatility
If diverse sensor data from various locations, perspectives, and lighting conditions is collected to ensure classifier robustness, then the adaptability and reliability of the classifier is improved, but the quantity of data required and the complexity of data collection increases
Solution Approach 1:
The patent adds dimensional diversity to training data by systematically varying parameters such as virtual sensor position, orientation, lighting conditions, and scene modifications. Instead of collecting diverse data in the physical world, the solution creates diversity by exploring additional parameter spaces through computational generation, achieving classifier robustness without increasing physical data collection complexity.
Solution Approach 2:
A single labeled scene serves multiple functions as the source for generating all derivative training samples. The base scene is reused and transformed through various operations (virtual sensor placement, lighting changes, scene modifications) to produce diverse training data, eliminating the need for multiple separate data collection campaigns and simplifying the overall process.
3Quantity of substance
If a large number of images are collected and labeled by hand to sufficiently train a classifier, then the quantity of training data is improved, but the cost and time investment increases significantly
Solution Approach 1:
The patent segments the training data generation process into two distinct phases: (1) manual labeling of a single base scene, and (2) automated generation of numerous derivative samples through virtual sensor placement and scene transformation. This segmentation allows the time-consuming labeling task to be performed once on a small subset, while the remaining quantity requirements are met through efficient automated processing.
Solution Approach 2:
The patent generates large quantities of training data by creating synthetic copies through virtual sensor captures and scene derivations. Each derivative image is an automatically generated copy based on transformations of the original labeled scene, producing abundant training data without proportional increases in manual labeling costs or time investment.
Data Source
AI summary
Data representing a scene is received. The scene includes labeled elements such as walls, a floor, a ceiling, and objects placed at various locations in the scene. The original received scene may be modified in different ways to create new scenes that are based on the original scene. These modifications include adding clutter to the scene, moving one or more elements of the scene, swapping one or more elements of the scene with different labeled elements, changing the size, color, or materials associated with one or more of the elements of the scene, and changing the lighting used in the scene. Each new scene may be used to generate labeled training data for a classifier by placing a virtual sensor (e.g., a camera) in the new scene, and generating sensor output data for the virtual sensor based on its placement in the new scene.