Auto-Labeling Perception Models for Autonomous Driving Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and inefficiency of manually annotating training data for machine learning models in autonomous driving systems, particularly for spatiotemporal models, hinder the development of accurate perception systems.
Innovation Solution
A two-stage training approach using a first model to auto-label data for a larger second model, which is then fine-tuned to perform the main perception task, enabling the generation of more accurate auto-labels and reducing human involvement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is used to create training data for machine learning models, then annotation quality and accuracy are improved, but annotation cost and time consumption increase significantly
Solution Approach 1:
The system performs preliminary action by training an auto-labeling model on initially manually annotated data before actual production use. This pre-trained model then automatically annotates subsequent training datasets, eliminating the need for repeated manual annotation while maintaining quality standards established during the preliminary training phase.
Solution Approach 2:
The system implements self-service by enabling the machine learning model to automatically generate its own training annotations through the auto-labeling process. The model uses its learned perception capabilities to annotate data without human intervention, making the annotation process self-sustaining after initial setup.
2Measurement precision
If manual annotation is used to create training data for machine learning models, then annotation quality and accuracy are improved, but labor cost increases significantly
Solution Approach 1:
The system implements self-service by enabling the machine learning model to automatically generate its own training annotations through the auto-labeling process. The model uses its learned perception capabilities to annotate data without human intervention, making the annotation process self-sustaining after initial setup.
Solution Approach 2:
The system uses copying by having the auto-labeling model replicate the annotation quality of manual processes. The model learns from high-quality manually annotated data and copies that annotation style and accuracy when generating automatic labels, maintaining quality while eliminating ongoing manual labor costs.
3Extent of automation
If a first machine learning model is used to generate annotations for training data, then automation level increases, but annotation accuracy decreases
Solution Approach 1:
The system performs preliminary action by training the auto-labeling model on a dataset that has been manually annotated with high accuracy. This preliminary training phase allows the model to learn correct annotation patterns and standards before being deployed for automatic annotation, ensuring that automation maintains high accuracy levels from the start.
Solution Approach 2:
The system implements feedback by using the ground truth annotations from manually labeled datasets to evaluate and refine the auto-labeling model's performance. This feedback loop allows continuous improvement of annotation accuracy while maintaining high automation levels.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention at least relates to a computer-implemented method (100), as well as a computing device (200) thereof. The method (100) comprises obtaining (S102) a second dataset comprising a set of sensor data sequences, wherein each sensor data sequence comprises sensor data samples depicting a physical environment over a plurality of time instances, each sensor data sample having an associated annotation, generated by processing the sensor data sample through a first machine learning model being trained, using a first dataset, to perform a perception task, wherein the perception task comprises generating a prediction of a sensor data sample for a given time instance, given said sensor data sample as input; training (S104), using the second dataset, a second machine learning model to perform an augmented perception task, wherein the augmented perception task comprises generating a prediction of a sensor data sample for a time instance of a plurality of time instances of a sensor data sequence, given the remaining sensor data samples of said sensor data sequence as input; fine-tuning (S106), using a third dataset, the second machine learning model, to perform the perception task, wherein the third dataset comprises sensor data samples depicting a physical environment and that are annotated for the perception task; and providing (S108) the fine-tuned second machine learning model as a model for annotating training data for subsequent training of a production model, of an automated driving system, to perform the perception task.