Physics-Based Data Augmentation for Machine Learning Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning techniques face challenges in handling limited training data by generating additional data that adds noise and unrealistically modifies objects in videos and images, such as through rotation or sliding techniques, which are not based on realistic physics.
Innovation Solution
An embedding platform generates a machine learning model by augmenting objects with physical properties, using 3D models and physical property data to create realistic augmented data sequences, which are then used to train and optimize a second machine learning model for improved accuracy and precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional data augmentation techniques (rotation, sliding) are used to generate additional training data, then the quantity of training data is increased, but the data quality deteriorates due to added noise and unrealistic modifications
Solution Approach 1:
The patent changes the parameters of data augmentation from arbitrary geometric transformations to physics-based transformations by incorporating physical properties (mass, friction, gravity) into the augmentation process. This allows generating diverse training samples while maintaining physical realism, thus increasing data quantity without sacrificing quality
Solution Approach 2:
The patent introduces physical properties as an intermediary between the original data and augmented data. These physical properties serve as mediators that guide the transformation process, ensuring that augmented data follows realistic physical laws rather than arbitrary modifications, thereby maintaining data quality while expanding quantity
2Measurement precision
If more training data is generated to improve model accuracy, then the model performance is enhanced, but the complexity of the data processing pipeline increases
Solution Approach 1:
The patent merges the data augmentation process with physical property analysis by integrating multiple functions (data generation, physics simulation, and transformation) into a unified pipeline. This combination reduces overall system complexity while maintaining the ability to generate high-quality augmented data for improved model accuracy
3Adaptability or versatility
If arbitrary transformation functions are used for data augmentation, then the diversity of augmented data is increased, but the realism and applicability of the augmented data deteriorates
Solution Approach 1:
The patent transforms the augmentation approach by changing from arbitrary transformation parameters to physics-constrained parameters. By using physical properties (mass, friction, gravity) as constraints, the system generates diverse data transformations that remain physically realistic, thus maintaining both diversity and realism
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
A device receives images of a video stream, models for objects in the images, and physical property data for the objects, and maps the models and the physical property data to the objects in the images to generate augmented data sequences. The device applies different physical properties to the objects in the augmented data sequences to generate augmented data sequences with different applied physical properties, and trains a machine learning (ML) model based on the images to generate a first trained ML model. The device trains the ML model, based on the augmented data sequences with the different applied physical properties, to generate a second trained ML model, and compares the first trained ML model and the second trained ML model. The device determines whether the second trained ML model is optimized based on the comparison, and provides the second trained ML model when optimized.