Sensor Domain Adaptation via ML Data Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning programs trained on data from one sensor with specific parameters fail to detect objects accurately when deployed with sensors having different parameters, such as varying resolution, field of view, or pose, due to the need for extensive retraining with new data.
Innovation Solution
A system and method that determine a domain translation from the data space of a first sensor to that of a second sensor by using a machine learning program to transform annotation data, allowing the use of existing training data from the first sensor for the second sensor, including determining a rigid body transformation between sensors and generating transformed training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning program is trained on data from a first sensor with specific parameters, then the program can accurately detect objects in that data space, but the program fails to detect objects accurately when deployed with a second sensor having different parameters
Solution Approach 1:
The patent applies parameter changes by transforming the training data from the first sensor's data space to the second sensor's data space through domain translation. This involves adjusting and reinterpreting the data according to the different sensor parameters (resolution, field of view, pose) so that the machine learning program can maintain its detection accuracy when deployed on sensors with different specifications without requiring retraining
2Reliability
If training data is collected from a first sensor, then the machine learning program can be trained effectively, but collecting new training data from a second sensor with different parameters requires extensive time and resources
Solution Approach 1:
The patent uses copying by creating a transformed copy of the existing training data from the first sensor. Through domain translation, the system generates synthetic training data that represents the second sensor's data space using the first sensor's existing annotated data, thereby avoiding the need to collect and annotate entirely new data from the second sensor
Solution Approach 2:
The patent applies preliminary action by pre-establishing a domain translation model that maps the first sensor's data space to the second sensor's data space. This translation model is trained in advance using paired data from both sensors, allowing subsequent training data generation to be performed efficiently without repeating the complex data collection and annotation process
3Adaptability or versatility
If the machine learning program is retrained with new data from the second sensor, then the program can adapt to the new sensor parameters, but the retraining process requires large amounts of new annotated data
Solution Approach 1:
The patent generates synthetic training data by copying and transforming the existing first sensor data into the second sensor's data space representation. This allows the system to create sufficient training data for the second sensor without requiring actual captures from the second sensor, thereby reducing the quantity of new data that needs to be collected and annotated
4Adaptability or versatility
If different sensors with varying parameters are used in a vehicle, then the system can accommodate multiple sensor configurations, but the machine learning program cannot be efficiently reused across these different sensors
Solution Approach 1:
The patent achieves universality by creating a domain translation framework that enables a single machine learning program trained on one sensor to be universally applied to multiple sensors with different parameters. The translation model acts as a bridge that allows the same trained model to function effectively across different sensor configurations without requiring separate training for each sensor type
Data Source
AI summary
A system includes a computer programmed to receive first sensor data from a first sensor, wherein the first sensor data is defined in part by a first data space that includes first parameters of the first sensor, and second sensor data from a second sensor, wherein the second sensor data is defined in part by a second data space that includes second parameters of the second sensor, to input the first sensor data and the second sensor data to a machine learning program to train the machine learning program to determine a domain translation of data from the first data space to the second data space, and then to input a set of training data received from the first sensor to the trained machine learning program to generate a training dataset based on the determined domain translation of data from the first data space to the second data space.