Cross-Sensor Object Detector Training Using Annotation Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection systems in vehicles face challenges in training due to the time- and resource-intensive nature of data annotation, particularly when transitioning between different sensor types like radar and LIDAR, leading to reduced detection quality.
Innovation Solution
A method and control device that utilize a sensor model, such as an artificial neural network, to transfer annotations from one sensor type to another, allowing for training on unlabeled data and optimizing the measurement characteristics of the target sensor, thereby enhancing the reusability of existing data records and reducing the need for additional annotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If annotations are created for one sensor type (e.g., LIDAR), then the annotation work is time- and resource-intensive, but directly using these annotations for another sensor type (e.g., radar) leads to reduced detection quality due to different measuring principles and sensor specifications
Solution Approach 1:
The patent introduces a sensor model as an intermediary that translates annotations from one sensor type to another. The sensor model learns the relationship between different sensor types and uses this knowledge to generate synthetic annotations for the target sensor, thereby avoiding direct annotation while maintaining detection quality.
Solution Approach 2:
The patent creates synthetic copies of annotations by using the sensor model to generate pseudo-labels for the target sensor data. Instead of directly copying annotations from the source sensor, the model transforms them into sensor-specific representations that account for the target sensor's measurement characteristics.
2Measurement precision
If annotations are created for every sensor type, then detection quality is maintained, but the time and resource cost increases significantly
Solution Approach 1:
The patent generates synthetic annotations by copying and transforming annotations from source sensor data through the sensor model. This allows the system to create training data for the target sensor without manual annotation, significantly reducing time and resource requirements while maintaining detection quality.
Solution Approach 2:
The sensor model enables the system to self-generate annotations for the target sensor by learning the sensor-specific characteristics and transforming source annotations accordingly. This self-service mechanism eliminates the need for manual annotation of every sensor type.
3Measurement precision
If sensor-specific annotations are used, then detection quality is optimized for that sensor type, but the adaptability to other sensor types is reduced
Solution Approach 1:
The sensor model serves as a universal translator that can handle multiple sensor types. By learning the characteristics of different sensors and the relationships between them, the model enables annotations to be adapted across various sensor types, making the system versatile and broadly applicable.
Solution Approach 2:
The sensor model acts as a mediator that decouples the annotation process from sensor-specific constraints. It translates source annotations into target-sensor-appropriate representations, enabling cross-sensor adaptation while maintaining sensor-specific detection accuracy.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach improves the quality and efficiency of object detection by increasing the usable data volume and reducing costs, enabling effective training across different sensor types without relying on direct annotations, thus enhancing the performance of object detectors.
Implementation Method 1
a sensor model, in particular an artificial neural network, is designed to image object representations onto sensor data of the sensor
Data Source
AI summary
A method is for training an object detector configured to detect objects in sensor data of a sensor. The method includes providing first sensor data of the sensor, providing an object representation assigned to the first sensor data, and transmitting the object representation to a sensor model. The method further includes imaging object representations onto the first sensor data of the sensor with the sensor model, assigning the object representation to second sensor data with the sensor model, and training the object detector based on the second sensor data.