Semi-supervised Object Detection Model Update via Voxel Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In autonomous driving environments, deep-learning-based object detectors face a trade-off between real-time performance and accuracy, and there is a need for continuous model updates using labeled and unlabeled voxel data while minimizing labeling costs and efforts.
Innovation Solution
A semi-supervised learning method for object detection that receives unlabeled voxel data from vehicles, updates the object detection model using labeled and unlabeled data, and incorporates a loss-based update mechanism involving supervised and consistency losses, with data and loss-based updates, and includes deidentification of voxel data for privacy protection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep-learning-based object detectors are used in autonomous driving environments, then detection accuracy is improved, but real-time performance deteriorates due to the trade-off relationship between accuracy and real-time characteristics
Solution Approach 1:
The patent segments the training data into labeled voxel data and unlabeled voxel data, and segments the loss function into supervised loss (from labeled data) and consistency loss (from unlabeled data). This allows the system to process different types of data through different learning mechanisms simultaneously, improving overall detection accuracy while maintaining real-time performance through efficient use of available data
Solution Approach 2:
The patent changes the learning parameters by introducing a composite loss function that combines supervised loss and consistency loss with adjustable weighting. This allows dynamic adjustment of learning priorities between accuracy and speed, enabling the system to adapt to different real-time performance requirements while maintaining high detection accuracy
2Productivity
If continuous model updates are performed using labeled voxel data, then learning efficiency is improved, but labeling costs and efforts increase
Solution Approach 1:
The system performs self-service by using unlabeled voxel data from autonomous vehicles to generate consistency loss, which automatically provides learning signals without requiring manual labeling. The semi-supervised learning framework enables the model to learn from its own predictions on unlabeled data, reducing dependency on expensive labeled data while maintaining learning efficiency
Solution Approach 2:
The patent merges labeled and unlabeled voxel data into a unified training framework, combining supervised learning (from labeled data) and self-supervised learning (from unlabeled data). This integration allows the system to leverage both types of data simultaneously, improving learning efficiency while minimizing the proportion of expensive labeled data required
3Measurement precision
If voxel data is collected from vehicles for learning, then model accuracy is improved, but personal information protection becomes compromised
Solution Approach 1:
The patent extracts only the essential geometric and spatial features from voxel data that are necessary for object detection, while removing or anonymizing personal information such as vehicle identifiers, location metadata, and other sensitive attributes. This extraction process retains the useful information for training while eliminating privacy risks
Data Source
AI summary
A semi-supervised learning method for object detection in an autonomous vehicle and a device for performing semi-supervised learning for object detection in an autonomous vehicle can include receiving, by a server, no-label voxel data from a vehicle, performing, by the server, a data-based update on a server object detection model on the basis of label voxel data and the no-label voxel data, determining, by the server, a loss value on the basis of the label voxel data and the no-label voxel data, and performing, by the server, a loss-based update on the server object detection model using the loss value.