AI Training Data Generation Using Hybrid Image Capture and Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for generating training data for AI applications are either time-consuming and expensive due to manual labeling or lack realism without CAD models, and synthetic approaches fail to replicate real-world scenarios accurately.
Innovation Solution
A hybrid method that captures real-world images from multiple perspectives, using operator input and neural networks to generate high-quality, error-free labels without requiring CAD models, by combining RGB and depth images and employing automated capture and verification processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling by human experts is used, then data accuracy and real-world correspondence are improved, but time consumption and cost increase
Solution Approach 1:
The labeling process is segmented into multiple stages: automated candidate region generation from 3D models, operator verification/correction, and final label generation. This divides the originally monolithic manual labeling task into manageable segments that leverage both automated efficiency and human expertise where needed.
Solution Approach 2:
3D CAD models are used to pre-generate candidate object regions and labels before human operators review them. This preliminary automated labeling reduces the workload on operators from creating labels from scratch to merely verifying and correcting pre-generated candidates, significantly reducing time consumption while maintaining accuracy.
2Productivity
If synthetic data generation using photorealistic 3D rendering is used, then time consumption is reduced, but realism and correspondence to real-world scenarios deteriorate
Solution Approach 1:
Real captured images serve as an intermediary between synthetic 3D models and final training data. The system captures real images of objects, uses them to train neural networks that then generate labels for synthetic 3D rendered images. This intermediary real-world data ensures the labels correspond accurately to real-world scenarios while maintaining high productivity through automated processing.
Solution Approach 2:
The patent replaces manual mechanical labeling operations with automated neural network-based labeling systems trained on real-world data. This substitution maintains high realism by learning from actual images while achieving automated productivity, eliminating the need for time-consuming manual labeling of each individual image.
3Productivity
If CAD models are required for synthetic data generation, then data generation efficiency is improved, but complexity of the process increases due to CAD file requirements
Solution Approach 1:
Instead of requiring complex CAD files and 3D modeling processes, the system creates simplified digital representations or point-cloud models directly from captured real-world images. This copying approach maintains the essential geometric information needed for labeling while eliminating the complexity of CAD file creation and management, reducing process complexity while preserving data generation efficiency.
Data Source
AI summary
The invention relates to a method for generating training data (12), comprising the steps of: - capturing at least one first image (26a, 26b) and a second image (28a, 28b) containing the object (14); - capturing an input (30) from an operator (32) regarding the position of the object (14) in the displayed first image (26a, 26b); - determining the object (14) in the first image (26a, 26b) based on the input (30); - generating first object information (34) based on the determined object (14); - determining the object (14) in the second image (28a, 28b) based on the determined first object information (34); - generating second object information (36) based on the determined object (14); and - generating training data (12) for the object (14). Furthermore, the invention relates to a computer program product, a computer-readable storage medium and a training system (10).