Augmented Pseudo-Labeling for Object Detection Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current state-of-the-art object detectors in autonomous systems face challenges in providing reliable perception measurements due to incomplete training data, leading to imperfect object detection, which is mitigated by deploying additional sensors like LIDAR or multi-camera systems at increased costs.
Innovation Solution
The implementation of augmented pseudo-labeling techniques for object detection learning, where unlabeled images are processed to generate domain-specific augmented images with associated labels, allowing for the retraining of machine learning models to improve detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If additional sensors like LIDAR or multi-camera systems are deployed to mitigate imperfect object detection, then reliability of perception measurements is improved, but system cost increases
Solution Approach 1:
The patent creates pseudo-labels by copying and transforming existing labeled data through domain-specific augmentations. Instead of collecting new physical sensor data, the system generates synthetic training examples that replicate real-world variations, thereby improving model reliability without adding expensive hardware sensors
Solution Approach 2:
The system modifies training data parameters by applying domain-specific augmentations (scaling, rotation, color adjustments, etc.) to existing images. This transforms the training data distribution to better match target domain conditions, improving detection reliability across varying environmental conditions without additional sensors
2Device complexity
If machine learning models are trained with limited training data, then device complexity is reduced, but detection accuracy deteriorates
Solution Approach 1:
The system performs preliminary data preparation by generating pseudo-labeled training examples before actual model training. This pre-processing step expands the effective training data size and improves data quality through domain-specific augmentations, enabling higher detection accuracy without requiring proportionally more raw training data
Solution Approach 2:
The patent introduces pseudo-labels as an intermediary between raw unlabeled data and final model training. These pseudo-labels serve as a bridge that transforms limited raw data into enriched training examples, improving detection accuracy while avoiding the need for extensive annotated training data
Data Source
AI summary
A method includes obtaining an image of a scene and identifying one or more labels for one or more objects captured in the image. The method also includes generating one or more domain-specific augmented images by modifying the image, where the one or more domain-specific augmented images are associated with the one or more labels. In addition, the method includes training or retraining a machine learning model using the one or more domain-specific augmented images and the one or more labels. Generating the one or more domain-specific augmented images may include at least one of modifying the image to include a different amount of motion blur, modifying the image to include a different lighting condition, and modifying the image to include a different weather condition.