Server Analyzes Meta Information to Target Vehicle Data Capture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for building training datasets for autonomous driving are inadequate as they require vast amounts of data, struggle with data redundancy, and lack effective meta-data compositionality, leading to suboptimal machine learning models with biases and inefficient data storage.
Innovation Solution
A method that analyzes meta-information of training datasets to determine requirements for extension, using vector representations to identify gaps and send targeted data capturing tasks to vehicles equipped with sensors like LIDAR and cameras, ensuring diverse data collection and avoiding redundancy by clustering and compensating for imbalanced clusters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If vast amounts of sensor data are collected from a fleet of vehicles, then the training dataset size increases, but storage costs and data handling complexity increase significantly
Solution Approach 1:
The system performs preliminary analysis of meta information (vector representations) before collecting actual sensor data. The server analyzes the vector space to identify gaps and determines data capturing tasks in advance, so that only necessary data is collected later by vehicles, avoiding unnecessary data storage and handling
Solution Approach 2:
The system extracts and analyzes only the essential meta information (vector representations) from sensor data, separating this analysis from the full sensor data collection. This allows the system to determine data requirements without storing or processing all raw sensor data, reducing storage and handling complexity
2Adaptability or versatility
If more sensor data is collected to cover different scenarios, then the diversity of training data improves, but redundant datapoints increase storage requirements
Solution Approach 1:
The system uses feedback from analyzing the vector representations of existing training data to identify gaps in scenario coverage. The server continuously monitors the training dataset composition and sends targeted data capturing tasks to vehicles only for scenarios that are underrepresented, avoiding collection of redundant data
Solution Approach 2:
The system changes the parameter of data selection from random or uniform collection to targeted collection based on vector space analysis. By transforming the problem into vector space and analyzing distances between vectors, the system identifies which scenarios need more data and adjusts the data capturing tasks accordingly
3Quantity of substance
If the training dataset is extended without targeted selection, then more datapoints are available, but the quality and usefulness for training decreases due to superfluous datapoints
Solution Approach 1:
The system transforms the data selection problem into vector space analysis, where each datapoint is represented as a vector. By analyzing the distribution and gaps in this vector space, the system can identify which types of scenarios are missing and target data collection to fill those gaps, ensuring high training quality
Solution Approach 2:
The system performs preliminary analysis of the vector representations before collecting actual sensor data. This allows the system to pre-determine which scenarios need more data coverage and send targeted data capturing tasks, ensuring that collected datapoints will be useful for training
Data Source
AI summary
The present disclosure relates to a method for building a training dataset on a server, including the steps of: analyzing meta information of the training dataset for a requirement to extend the training dataset; and based on the requirement, sending a data capturing task to a data capturing device, in particular a vehicle.