Synthetic Depth Map Training for Autonomous Vehicle Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems used for object recognition in autonomous vehicles face challenges in reliably recognizing objects due to the complexity of scenarios, especially when image data and depth information may deviate, leading to insufficient training on rare or unusual cases like optical illusions.
Innovation Solution
The method involves training machine learning systems using synthetically generated depth maps in addition to image data, allowing for the adaptation of parameter values to improve object recognition by simulating diverse scenarios and enhancing the system's ability to handle ambiguous and rare cases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning systems are trained only with real image data and real depth maps, then the training data collection is time-consuming and expensive, but the system reliability is limited due to insufficient coverage of rare cases
Solution Approach 1:
The patent uses synthetic depth maps as artificial copies of real depth data to train the machine learning system. These synthetic depth maps are generated from image data through simulated depth estimation, creating realistic training examples without requiring actual sensor measurements. This copying approach allows extensive training data generation quickly and inexpensively while maintaining sufficient realism to improve system reliability on rare cases
Solution Approach 2:
The patent performs preliminary synthesis of depth maps before actual training occurs. By pre-generating synthetic depth maps with known ground truth labels and varying scenarios (including rare cases), the system prepares comprehensive training data in advance. This preliminary action eliminates the need for time-consuming real-world data collection and annotation for each training scenario
2Reliability
If machine learning systems use multipath processing with both image data and depth information, then object recognition is improved, but the complexity of training data acquisition and processing increases
Solution Approach 1:
The patent merges the training of image processing and depth estimation into a unified machine learning system. Instead of separately training image recognition and depth map generation, the system jointly processes both data types through integrated neural network architectures. This merging reduces training complexity by consolidating multiple training procedures into a single coordinated process while maintaining the benefits of multipath processing
3Productivity
If synthetic depth maps are generated and used for training, then training efficiency is improved and coverage of rare cases is enhanced, but the manufacturing precision of depth information may be reduced
Solution Approach 1:
The patent applies parameter changes by adjusting synthesis parameters such as noise levels, depth estimation algorithms, and scene configurations during synthetic depth map generation. By varying these parameters to match real-world conditions and uncertainty characteristics, the synthetic data maintains sufficient precision for training while enabling efficient generation of diverse scenarios including rare cases that would be difficult to capture in real data
Data Source
AI summary
A method for training a machine learning system, in which image data are fed into a machine learning system with processing of at least a part of the image data by the machine learning system. The method includes synthetic generation of at least a part of at least one depth map that includes a plurality of depth information values. The at least one depth map is fed into the machine learning system with processing of at least a part of the depth information values of the at least one depth map. The machine learning system is then trained based on the processed image data and based on the processed depth information values of the at least one depth map, with adaptation of a parameter value of at least one parameter of the machine learning system, the adapted parameter value influencing an interpretation of input data by the machine learning system.