Autonomous Vehicle Labeling Model Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face inefficiencies due to manual labor requirements in generating and reviewing bounding boxes from sensor data, which are prone to inaccuracies leading to displaced or jittering object trajectories.
Innovation Solution
A system that uses a first labeling model to generate bounding boxes, with user adjustments to correct inaccuracies, and trains a second labeling model to improve accuracy by learning from these adjustments, reducing manual effort and enhancing object tracking and trajectory smoothing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to generate and review bounding boxes, then accuracy of object detection can be improved, but time consumption and cost increase significantly
Solution Approach 1:
The system performs preliminary automated bounding box generation using a first labeling model before manual review, preparing initial labels that reviewers can quickly correct rather than creating labels from scratch. This preliminary action reduces the time required for manual labeling while maintaining accuracy through subsequent review and correction steps.
Solution Approach 2:
The system implements feedback by using user corrections to retrain and improve the labeling model. The corrected bounding boxes from manual review are fed back into the system to train a second labeling model, which learns from these corrections and progressively improves accuracy, reducing the need for extensive manual labeling in future iterations.
2Productivity
If automated labeling models are used to generate bounding boxes, then productivity increases, but measurement precision decreases due to sensing inaccuracies
Solution Approach 1:
The system introduces an intermediary human review step between automated labeling and final output. Users review and correct the bounding boxes generated by the first labeling model, serving as an intermediary that catches sensing inaccuracies. This hybrid approach maintains the productivity benefits of automation while improving precision through human oversight.
Solution Approach 2:
The system replaces the purely mechanical manual labeling process with an automated labeling model, substituting human manual drawing of bounding boxes with algorithmic generation. This substitution dramatically increases productivity while the subsequent human review step corrects the precision issues introduced by the automated system.
3Reliability
If multiple rounds of manual review are conducted to verify bounding boxes, then reliability of object detection improves, but device complexity and operational difficulty increase
Solution Approach 1:
The system performs preliminary automated generation of bounding boxes before manual review, so that reviewers are verifying pre-generated labels rather than creating them manually. This preliminary action simplifies the review process to primarily correction tasks, reducing operational complexity while maintaining reliability through the review step.
Solution Approach 2:
The system enables self-service by allowing the labeling model to automatically generate and initially verify its own outputs through iterative retraining. The second labeling model, trained on corrected data, progressively improves reliability autonomously, reducing the need for continuous manual review rounds and simplifying the overall process complexity.
Data Source
AI summary
Aspects of the disclosure relate to training a labeling model to automatically generate labels for objects detected in a vehicle's environment. In this regard, one or more computing devices may receive sensor data corresponding to a series of frames perceived by the vehicle, each frame being captured at a different time point during a trip of the vehicle. The computing devices may also receive bounding boxes generated by a first labeling model for objects detected in the series of frames. The computing devices may receive user inputs including an adjustment to at least one of the bounding boxes, the adjustment corrects a displacement of the at least one of the bounding boxes caused by a sensing inaccuracy. The computing devices may train a second labeling model using the sensor data, the bounding boxes, and the adjustment to increase accuracy of the second labeling model when automatically generating bounding boxes.