Training Data Annotation Using Multi-Time-Step State Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle systems face challenges in accurately estimating the states of dynamic entities due to noise and uncertainty in sensor data, particularly when ground truth data is scarce, and offline models that leverage future time steps are computationally prohibitive for real-time applications.
Innovation Solution
An offline model is used to generate pseudo-ground truth data by processing pointwise measurements from multiple time steps, incorporating a data association model and a smoother to refine estimates, which can then be used to benchmark and update a runtime model for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an offline model is used to process pointwise measurements from multiple time steps to generate pseudo-ground truth data, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The offline model performs preliminary processing of pointwise measurements from multiple time steps to generate pseudo-ground truth data before runtime use. This preliminary action resolves tracking ambiguities and refines state estimates in advance, improving measurement precision without adding real-time computational complexity to the runtime system.
Solution Approach 2:
The system creates a simplified copy of the runtime model (the offline model) that can be executed in advance to generate pseudo-ground truth data. This copy performs the complex processing offline, allowing the actual runtime model to remain computationally efficient for real-time operation while still benefiting from improved accuracy.
2Measurement precision
If an offline model processes data from multiple time steps including future time steps, then measurement precision is improved, but productivity decreases due to computational burden
Solution Approach 1:
The computationally intensive processing of multi-time step data is performed as a preliminary action offline, generating pseudo-ground truth data in advance. This separates the heavy computational burden from real-time operation, allowing accurate state estimation without compromising runtime productivity.
Solution Approach 2:
The processing is segmented into offline and runtime components. The offline model handles the complex multi-time step processing to generate pseudo-ground truth, while the runtime model uses this pre-processed data for efficient real-time operation. This segmentation allows high precision processing to occur outside the real-time constraint.
3Reliability
If ground truth data is used for training machine learning models, then reliability is improved, but loss of substance increases due to scarcity of ground truth data
Solution Approach 1:
The system creates pseudo-ground truth data as a copy substitute for actual ground truth data. By processing pointwise measurements through the offline model to generate pseudo-ground truth, the system creates sufficient training data without requiring scarce real ground truth annotations, enabling reliable model training.
Solution Approach 2:
The system uses its own offline model to generate pseudo-ground truth data that serves the training needs of machine learning models. This self-service approach creates training data internally from available pointwise measurements, eliminating dependence on externally obtained ground truth data.
Data Source
AI summary
A computer-implemented method. Includes determining an estimate of a state of an object detected at a first time step, based on a pointwise estimate of the state of the object at the first time step and pointwise measurements of the state of the object at a plurality of further time steps. Includes generating, using the estimate of the state of the object, a proposed annotation associated with the object at the first time step. Includes rendering, via a user interface, a visual representation of the environment at the first time step and a visual representation of the proposed annotation. Includes receiving, via the user interface, user input indicating a user-approved annotation associated with the object at the first time step. Includes generating training data for a machine learning model for use in controlling an autonomous vehicle, based at least in part on the user-approved annotation.