Two-Stage Active Learning for Automotive Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and complexity of generating large, labeled datasets for training machine learning algorithms for automotive perception tasks in vehicles, which are necessary for partial driving automation, are addressed by implementing a two-stage active learning approach that leverages vehicle and data center instances of an object classifier to selectively annotate and update the classifier using an oracle.
Innovation Solution
A two-stage active learning method where vehicles and a data center process sensor data using their respective instances of an object classifier, selecting data for annotation based on uncertainty and divergence criteria, and updating the classifier with annotated data to enhance accuracy for object classification tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large labeled datasets are generated manually to train machine learning algorithms for automotive perception tasks, then the accuracy of object classification is improved, but the cost and time required for data preparation increases significantly
Solution Approach 1:
The system performs self-service by automatically selecting unlabeled data that would be most beneficial to label based on uncertainty sampling and diversity criteria. The active learning framework enables the system to autonomously identify which data points should be sent to human annotators, reducing manual effort while maintaining high classification accuracy.
Solution Approach 2:
The system implements feedback loops where classification results are continuously evaluated, and uncertain predictions are fed back for manual annotation. This creates an iterative process where the model learns from targeted human feedback, improving accuracy over time with minimal labeling effort.
2Reliability
If more data is labeled manually to improve machine learning algorithm performance, then the reliability of driving automation features is improved, but the cost of data generation increases
Solution Approach 1:
The system extracts only the most valuable data points for manual labeling by applying uncertainty sampling and diversity selection criteria. Instead of labeling all available data, it extracts a small subset that maximizes model improvement, significantly reducing annotation costs while maintaining reliability.
Solution Approach 2:
The system changes the parameter of data selection from random or exhaustive labeling to intelligent selection based on uncertainty and diversity metrics. This parameter change enables the system to achieve the same reliability improvement with far fewer labeled examples.
3Device complexity
If a single-stage active learning approach is used to select data for annotation, then the process is simpler, but the accuracy improvement for driving automation tasks is insufficient
Solution Approach 1:
The active learning process is segmented into two distinct stages: (1) initial model training with a small labeled dataset, and (2) iterative uncertainty-based data selection and retraining. This segmentation allows each stage to be optimized independently, achieving high accuracy while managing complexity through structured progression.
Solution Approach 2:
The system performs preliminary action by training an initial model before the active learning loop begins. This preliminary model is then used to guide subsequent data selection, enabling the system to make informed decisions about which data to label first, thereby accelerating accuracy improvement.
Data Source
Figure 1~2
Figure 3A
Figure 3B
AI summary
The present disclosure relates to active learning to improve object classification by an object classifier, the object classification being used for one or more visual perception tasks enabling one or more driving automation system features. To this end, automotive sensor data is selected for annotation by an oracle based on a first selection of sets of automotive sensor data by one or more vehicles based on vehicle instances of the object classifier followed by a second selection by a data center based on repeating object classification on the first selection of automotive sensor data sets using a data center instance of the object classifier. The second selection of automotive sensor data is then annotated by the oracle and used to train the object classifier. Afterwards, both the vehicle instances and the data center instance of the object classifier are updated with the trained object classifier.