Automated Training Data Generation via Sensor Projection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The generation of training data for trainable image processing methods is labor-intensive and often requires high computational resources, especially when dealing with realistic and diverse image data.
Innovation Solution
A method that involves obtaining multiple detections with known relative ratios between sensors, determining and projecting content information across these detections, and checking for inconsistencies in the image content to automatically generate label data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to generate training data, then labeling accuracy is improved, but labor cost and time consumption increase significantly
Solution Approach 1:
The system performs preliminary actions by automatically generating candidate labels using detection algorithms and projection methods before manual verification. This preliminary labeling reduces the amount of manual work needed while maintaining accuracy, as annotators only need to verify and correct rather than create labels from scratch.
Solution Approach 2:
The system introduces an intermediary automated labeling process between complete manual labeling and no labeling. This intermediary system uses detection results from multiple sensors and projection algorithms to generate preliminary labels, which then serve as a foundation for manual verification, reducing overall time consumption while preserving accuracy.
2Measurement precision
If synthetic image data is used for training, then labeling accuracy is improved, but computational resources and time required for generation increase
Solution Approach 1:
Instead of generating entirely synthetic images with complex renderers, the system copies and projects real detection results from multiple sensors into target image spaces. This copying approach preserves the realism of actual sensor data while obtaining accurate labels, avoiding the enormous computational resources needed for photorealistic synthetic image generation.
Solution Approach 2:
The system changes the approach from generating images with complex physical parameters (lighting, materials, physics) to transforming detection results through geometric projection parameters. This parameter change from physical rendering to geometric transformation significantly reduces computational resource requirements while maintaining labeling accuracy.
3Productivity
If programmatic labeling is used to process more data, then productivity is improved, but handling inconsistencies between multiple detections becomes complex
Solution Approach 1:
The system segments the inconsistency handling process into distinct modules: detection result acquisition, label projection, consistency checking, and conflict resolution. By segmenting the complex task into manageable steps, the system can process multiple detections efficiently while systematically handling inconsistencies without overwhelming complexity.
Solution Approach 2:
The system implements feedback mechanisms where detection results from multiple sensors are projected and compared, with inconsistencies feeding back into the labeling process for resolution. This feedback loop allows the system to automatically identify and handle conflicts between detections, maintaining productivity while managing complexity through structured feedback processing.
Data Source
AI summary
A method for generating training data for a trainable method for a system including sensor(s) for detecting at least one subarea of the surroundings around the system. The method includes: a) obtaining first and second detections having at least one known relative ratio between the detections and/or the sensors that carried out the detections; b) determining a portion of the particular content of the detections, and assigning a piece of information concerning the determined content to the detection in question, c) projecting assigned piece of information from one of the detections and/or from a content representation associated with same into at least one other of the detections and/or into a content representation associated with the other detection, d) checking a subarea of at least one of the detections and/or of at least one of the content representations for possible inconsistencies in the detection content.