Cross-Sensor Ground Truth Weighting for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training artificial neural networks for tasks like classification or object detection requires large amounts of manually labeled data, which is costly and time-consuming, and inter-sensor perception discrepancies between sensors like radar and lidar complicate the training process.
Innovation Solution
A method for training machine-learning models using approximations of ground truths from a second sensor, such as lidar, while minimizing the impact of lower-quality approximations by employing an optimization criterion based on cross entropy and an energy map, and creating a reliability map to filter out unreliable data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manually labeled data is used for training, then training accuracy is improved, but training cost and time consumption increase
Solution Approach 1:
The patent uses approximations of ground truths from a second sensor (lidar) to create training labels for the first sensor (radar). Instead of manually labeling radar data, the system copies label information from lidar data through automated approximation processes, significantly reducing labeling time while maintaining acceptable training quality.
Solution Approach 2:
The patent introduces an intermediary approximation process that maps ground truths from one sensor domain to another. The second sensor's ground truths serve as an intermediary representation that can be transformed into training labels for the first sensor, avoiding direct manual annotation while preserving essential information.
2Productivity
If approximations of ground truths from a second sensor are used, then training efficiency is improved, but data quality and reliability deteriorate
Solution Approach 1:
The patent applies local quality by creating a reliability map that assigns different quality weights to different regions of the approximation data. High-quality regions that show good correspondence between sensors are weighted more heavily, while low-quality regions are downweighted or excluded, allowing the system to maximize the use of reliable approximation data.
Solution Approach 2:
The patent changes the parameter of data quality assessment by introducing a reliability metric that quantifies the quality of approximations. This parameter transformation allows the system to differentiate between high and low quality approximations and adjust training accordingly, converting unreliable data into a form that can be selectively used or discarded.
3Quantity of substance
If all approximation data is used for training, then data quantity is improved, but training quality deteriorates due to inclusion of low-quality data
Solution Approach 1:
The patent segments the approximation data into different quality categories using a reliability map. By dividing the data into high-quality and low-quality regions, the system can selectively use only the high-quality segments for training, maintaining training quality while still leveraging the large quantity of available approximation data.
Solution Approach 2:
The patent applies partial action by using only the portion of approximation data that meets quality thresholds. Rather than using all available data indiscriminately, the system selectively applies high-quality approximations, achieving sufficient training data quantity without the harmful effects of low-quality data.
Data Source
AI summary
A computer-implemented method for training a machine-learning method comprises the following steps carried out by computer hardware components: determining measurement data from a first sensor; determining approximations of ground truths based on a second sensor; and training the machine-learning method based on the measurement data and the approximations of ground truths; wherein approximations of ground truths of lower-approximation quality have a lower effect on the training than approximations of ground truths of higher-approximation quality.