Generating Ground Truth For Machine Learning From Time Series Elements

What is this patent about?

’461 is related to the field of machine learning, specifically techniques for improving the training of machine learning models used in autonomous driving systems. The background involves the challenge of creating high-quality training datasets for deep learning systems, which traditionally requires significant manual effort in collecting, curating, and annotating data. This is especially difficult for edge cases where the model needs the most improvement, creating a need for automated training data generation.

The underlying idea behind ’461 is to leverage a time series of sensor data to automatically generate ground truth data for training machine learning models. Instead of relying on single data points, the invention uses a sequence of sensor readings (e.g., camera images, lidar data) captured over time to build a more complete and accurate representation of the environment. This is achieved by combining the most reliable information from different points in the time series to create a composite ground truth.

The claims of ’461 focus on a method, computer program product, and system for training a machine learning model. The core process involves receiving sensor data as a group of time series elements, determining a ground truth for the entire group by identifying respective portions of individual time series elements, and then training the model using a dataset comprising this ground truth and a *single*, selected time series element. The trained model then predicts the ground truth based on that single element.

In practice, this allows the system to learn to predict complex environmental features, such as the three-dimensional trajectory of a lane line, from a single image. The system analyzes a video sequence of a vehicle driving along a lane, combining information from multiple frames to create a highly accurate 3D representation of the lane line. This composite 3D lane line then becomes the ground truth, which is paired with a single frame from the video to train the model. The model learns to infer the complete 3D lane line from just that one frame.

This approach differentiates itself from traditional methods by automating the ground truth generation process and improving the accuracy of the training data. Instead of manually labeling data, the system uses the time series to create a more robust and complete representation of the environment. This is particularly useful for features that are difficult to capture in a single frame, such as occluded objects or distant lane markings. By training on this automatically generated, high-quality data, the machine learning model can achieve improved performance and robustness in autonomous driving applications.