Sensor Data Signature Selection for Annotation Efficiency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine learning methods for automated perception and scene recognition require large amounts of manually annotated data, leading to high time and cost burdens due to the labor-intensive nature of scene annotation.

Innovation Solution

A computer-implemented method determines candidate data sets for labeling by analyzing sensor data sets to identify suitable data sets for annotation using signatures, which are numeric representations generated through neural networks or machine learning methods, allowing for efficient selection and reduction of data sets that need manual labeling.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If large amounts of sensor data sets are collected for machine learning, then the amount of training data increases, but the time and cost for manual annotation increases

Engineering Contradiction:
Improveamount of training dataVSAvoidtime for manual annotation
Core Design Contradiction:
Quantity of substanceVSLoss of time

Solution Approach 1:

The system performs preliminary actions by automatically generating signatures for sensor data sets before manual annotation is needed. These signatures pre-process and characterize the data, enabling rapid identification of suitable candidates for annotation without requiring manual inspection of every data set, thus reducing annotation time while maintaining adequate training data quantity

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The signature acts as an intermediary between the raw sensor data sets and the manual annotation process. Instead of annotators directly reviewing large volumes of raw data, the signature serves as a compressed representation that mediates the selection process, allowing efficient filtering and identification of annotation candidates while reducing overall annotation time

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If manual annotation is performed on all sensor data sets, then the completeness of labelled data increases, but the cost and time consumption increases

Engineering Contradiction:
Improvecompleteness of labelled dataVSAvoidannotation efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The system applies local quality by not treating all sensor data sets uniformly for annotation. Instead, it identifies and prioritizes specific local regions or characteristics within the data (through signature analysis) that are most valuable for training, focusing annotation resources on data sets with high annotation value rather than uniformly processing all data

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system changes the parameter representation of sensor data sets from raw form to signature form. This parameter transformation enables efficient comparison and selection of data sets for annotation, allowing the system to maintain reliability by selecting high-quality data while improving productivity through rapid signature-based filtering and selection

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If signatures are generated for all sensor data sets, then the ability to identify similar scenes improves, but the computational complexity increases

Engineering Contradiction:
Improvesimilarity identification accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system extracts the essential characteristics from sensor data sets by generating signatures that capture the most important features for scene similarity identification. This extraction process condenses complex raw data into compact signature representations, maintaining measurement precision for similarity identification while reducing the computational complexity of subsequent comparison and selection operations

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3985560A1Methods and systems for determining candidate data sets for labelling
Publication Date: 2022.04.20 APTIV TECHNOLOGIES AG
  • EP3985560A1 patent drawingFigure 1
  • EP3985560A1 patent drawingFigure 2
  • EP3985560A1 patent drawingFigure 3

AI summary

A computer implemented method for determining candidate data sets for labelling comprises the following steps carried out by computer hardware components: determining a plurality of sensor data sets; determining a respective signature for each of the plurality of sensor data sets; determining, based on the signature of the respective sensor data set, for each of the plurality of sensor data sets whether the respective sensor data set is a candidate data set for labelling; and providing the sensor data set to a labeling instance for labelling if the sensor data set is a candidate data set for labelling.