Training Data Generation with Multi-Value Region Labels

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing region estimation systems face challenges in accurately utilizing the knowledge of multiple annotators due to the conversion of multi-value annotations into binary labels, leading to information loss and reduced detection accuracy.

Innovation Solution

A training data generation device that assigns labels from three or more different values based on the designated regions by multiple annotators, preserving the nuanced likelihood information, and generates training data for a region estimator.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If multiple annotators' designated regions are converted into binary labels (presence/absence), then the training data structure becomes simple and easy to process, but information about the degree of likelihood is lost and detection accuracy deteriorates

Engineering Contradiction:
Improveease of training data processingVSAvoidloss of likelihood information
Core Design Contradiction:
Ease of manufactureVSLoss of information

Solution Approach 1:

The patent changes the label parameter from binary (0/1) to multi-valued (0, 1, 2, 3) to represent different degrees of likelihood. This allows the training data to preserve the nuanced information from multiple annotators while maintaining a structured format suitable for processing. The label assignment unit assigns these multi-valued labels based on the number of annotators designating each region, thereby resolving the contradiction between simplicity and information preservation.

Inventive Principle:
Principle #35Parameter changes

2Ease of manufacture

If multiple annotators' designated regions are converted into binary labels, then the label structure becomes simple, but detection accuracy of the region estimator deteriorates

Engineering Contradiction:
Improveease of label structureVSAvoiddetection accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent modifies the label parameter from binary to multi-valued to capture the degree of likelihood, directly improving detection accuracy while maintaining structural simplicity.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The label assignment unit acts as an intermediary that processes multiple annotators' designations and converts them into multi-valued labels. This intermediary step preserves the collective knowledge of annotators and translates it into a format that the region estimator can effectively learn from, thereby improving detection accuracy without complicating the overall system.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Device complexity

If knowledge from multiple annotators is not fully utilized, then the processing complexity is reduced, but the detection accuracy and precision of region estimation deteriorates

Engineering Contradiction:
Improveprocessing complexityVSAvoidregion estimation precision
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

By changing the label parameter to multi-valued, the patent enables the system to incorporate knowledge from multiple annotators in a straightforward manner. The label assignment unit automatically processes annotators' designations and assigns appropriate multi-valued labels, improving region estimation precision without significantly increasing processing complexity.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4621727A1Training data generation device, learning device, region detection device and computer program
Publication Date: 2025.09.24 CASIO COMPUTER CO LTD
  • EP4621727A1 patent drawingFigure 1~2
  • EP4621727A1 patent drawingFigure 3~4
  • EP4621727A1 patent drawingFigure 5~6

AI summary

A training data generation device includes one or more processors configured to: obtain a plurality of designated regions designated by a plurality of annotators as detection target regions with respect to a target image; assign, as labels, data selected from three or more different values indicating the degree of likelihood of being a detection target to respective regions in the target image based on the plurality of designated regions; and generate training data for a region estimator that estimates a detection target region by associating the target image with the labels assigned to the respective regions of the target image.