3D Simulator for Synthetic Machine Learning Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine-learning-based object recognition systems for autonomous vehicles face challenges in cost and performance due to the need for large volumes of manually created image data and teacher data, particularly in diverse environments like mines, leading to potential overlearning and reduced detection accuracy when data is scarce.
Innovation Solution
A database construction system that automatically generates virtual three-dimensional shape data and corresponding teacher data using a three-dimensional simulator, enabling the creation of large volumes of learning information without manual input, specifically utilizing unmanned aerial vehicles for data acquisition and processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large volumes of image data and teacher data are manually collected and created for machine learning, then detection accuracy can be improved, but costs and time consumption increase significantly
Solution Approach 1:
The patent uses a three-dimensional simulator to generate virtual image data and teacher data as copies of real-world scenarios. Instead of manually collecting actual images, the system creates synthetic training data that replicates various driving conditions, obstacles, and environments, thereby reducing data collection time while maintaining detection accuracy.
Solution Approach 2:
The system performs preliminary data generation by creating comprehensive virtual training datasets before actual machine learning training begins. The three-dimensional simulator pre-generates diverse scenarios including different weather conditions, lighting, and obstacle configurations, so that when training starts, the model already has extensive pre-prepared data to learn from.
2Adaptability or versatility
If diverse image data from multiple environments is collected to prevent overlearning, then detection robustness improves, but data collection costs increase
Solution Approach 1:
The three-dimensional simulator creates virtual copies of diverse environmental conditions including different weather scenarios, lighting conditions, and geographical locations. This allows the system to generate unlimited diverse training data without the physical costs of traveling to and collecting data from multiple real-world locations.
Solution Approach 2:
The virtual simulation system serves multiple functions simultaneously: it generates image data, creates corresponding teacher data, varies environmental conditions, and simulates different sensor perspectives. This multi-functional approach replaces multiple separate data collection processes, reducing overall costs while maintaining data diversity.
3Manufacturing precision
If manual creation of teacher data is performed sheet by sheet for each image, then detection precision improves, but manufacturing complexity and costs increase
Solution Approach 1:
The system automatically generates teacher data as corresponding copies to the virtual image data. Since both the images and their teacher labels are created through the same three-dimensional simulation process, the pairing is automatic and precise, eliminating the manual sheet-by-sheet creation process while maintaining high accuracy.
Solution Approach 2:
The three-dimensional simulator performs self-service by automatically generating both image data and corresponding teacher data without external human intervention. The system autonomously creates consistent paired datasets, reducing the need for manual data annotation processes and associated complexity.
Data Source
AI summary
An object is to provide a database construction system for machine-learning that can automatically simply create virtual image data and teacher data in large volumes. A database construction system for machine-learning includes: a three-dimensional shape data input unit configured to input three-dimensional shape information about a topographic feature or a building acquired at three-dimensional shape information measuring means; a three-dimensional simulator unit configured to automatically recognize and sort environment information from the three-dimensional shape information; and a teacher data output unit configured to output virtual sensor data and teacher data based on the environment information recognized at the three-dimensional simulator unit and a sensor parameter of a sensor.