Learning Data Selection Using Sensor Agreement in Object Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods face challenges in obtaining learning data for machine learning models when using sensors with low precision, as they often reject detection results with low confidence, limiting the availability of training data and increasing processing loads.

Innovation Solution

An information processing method that acquires object detection results from both high and low precision sensors, determines the degree of agreement between them in specific regions, and selects sensing data as learning data based on this agreement, ensuring effective learning data acquisition even with low precision sensors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If detection results with low precision are rejected as learning data, then the precision of training data is improved, but the quantity of available learning data is reduced

Engineering Contradiction:

Improveprecision of training dataVSAvoidquantity of learning data

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent applies local quality by evaluating detection results differently based on their spatial location. Detection results within a predetermined region (e.g., regions of interest) are selected as learning data even if they have low precision, while other results may be rejected. This allows the system to maintain high data quality in critical areas while preserving quantity through selective acceptance based on location rather than uniform precision thresholds

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent changes the selection parameter from precision alone to a combination of precision and spatial location. By introducing location as an additional parameter, the system can accept low-precision detection results when they occur in predetermined regions, thereby increasing the quantity of learning data without compromising the overall quality, as the location-based filtering ensures only relevant data is included

Inventive Principle:

Principle #35Parameter changes

2Quantity of substance

If all detection results are used as learning data regardless of precision, then the quantity of learning data is increased, but the precision of training data deteriorates

Engineering Contradiction:

Improvequantity of learning dataVSAvoidprecision of training data

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent applies local quality by implementing location-based filtering where only detection results within predetermined regions are selected as learning data. This spatial filtering mechanism ensures that even though the system processes large quantities of detection results, only those in relevant locations are used for training, thereby maintaining high precision in the training data while still achieving sufficient quantity through the selective acceptance of low-precision results in specific regions

Inventive Principle:

Principle #3Local quality

3Reliability

If learning data is selected based on high precision thresholds, then data quality is improved, but the processing load increases due to extensive filtering

Engineering Contradiction:

Improvequality of learning dataVSAvoidprocessing load

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by pre-defining regions of interest before the learning data selection process. These predetermined regions are established in advance based on expected object locations or areas of interest. During data selection, the system only needs to check whether detection results fall within these pre-defined regions, rather than performing complex precision analysis on all results. This preliminary preparation significantly reduces processing load while maintaining data quality, as the filtering criterion becomes a simple spatial check rather than a complex precision evaluation

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11915099B2Information processing method, information processing apparatus, and recording medium for selecting sensing data serving as learning data

Publication Date: 2024.02.27 PANASONIC INTELLECTUAL PROPERTY CORP OF AMERICA

AI summary

An information processing method includes: acquiring a first object detection result obtained by use of an object detection model to which sensing data from a first sensor is input, and a second object detection result obtained by use of a second sensor; determining a degree of agreement between the first object detection result and the second object detection result in a specific region in a sensing space of the first sensor and the second sensor; and selecting the sensing data as learning data for the object detection model, according to the degree of agreement obtained in the determining.