Object Detection Class Aggregation for Fine-to-Coarse Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection technologies face challenges in improving detection performance for applications requiring detection of objects in rough classes while maintaining accuracy for fine multiple classes, particularly when common characteristics among classes complicate identification.

Innovation Solution

An object detection system utilizing a first class definition with fine multiple classes for precise object position region detection and class identification, followed by a second class definition with fewer classes for output, enhancing detection performance by associating the two definitions to balance precision and simplicity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If object position region detection unit and class identification unit learn using data with few classes, then it is possible to detect objects regarding the few classes and output detection result, but detection performance of object at the time of inference is deteriorated

Engineering Contradiction:
Improvenumber of classes in learning dataVSAvoiddetection performance
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the detection task into two stages: first detecting objects with fine-grained multiple classes to capture detailed characteristics, then grouping these fine classes into coarse classes for the final output. This segmentation allows the system to learn from rich fine-class data while providing simplified coarse-class results, resolving the contradiction between using few classes and maintaining detection performance.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If object position region detection unit and class identification unit learn using data with fine multiple classes, then it becomes possible to detect objects in multiple classes at the time of inference, but it is not effective for applications not requiring detection result of objects in fine multiple classes

Engineering Contradiction:
Improvedetection performanceVSAvoidnumber of classes in output
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements a dynamic class aggregation mechanism where fine classes detected during inference are dynamically grouped into coarse classes based on predefined mappings. This allows the system to adapt its output granularity according to application needs, providing detailed fine-class results when necessary and simplified coarse-class results for applications like traffic census, thereby resolving the contradiction between detection performance and output complexity.

Inventive Principle:
Principle #15Dynamics

3Ease of operation

If applications require only counting of rough classes such as person, vehicle, and animal, then it is not necessary to distinguish fine classes such as car, bus, and truck, but learning with few classes deteriorates detection performance

Engineering Contradiction:
Improvesimplicity of detection resultVSAvoiddetection performance
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent performs preliminary detection with fine classes first, capturing comprehensive object characteristics, then applies post-processing aggregation to group fine classes into coarse classes based on application requirements. This preliminary action with fine classes ensures high detection performance is achieved during learning, while the subsequent aggregation step provides the simplified coarse-class output needed for applications like traffic census, resolving the contradiction between simplicity and performance.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3975112B1Object detection device, object detection method, program, and recording medium
Publication Date: 2025.07.09 KONICA MINOLTA INC
  • EP3975112B1 patent drawingFigure 1
  • EP3975112B1 patent drawingFigure 2
  • EP3975112B1 patent drawingFigure 3

AI summary

An object position region detection unit of an object detection device detects a position region of an object included in an input image on the basis of a first class definition in which a plurality of classes is defined in advance. A class identification unit identifies a class out of a plurality of classes to which the object belongs on the basis of the first class definition. An object detection result output unit outputs class information of the object as a detection result of the object on the basis of a second class definition in which a plurality of classes is defined in advance, the second class definition associated with the first class definition. The number of classes defined in the second class definition is smaller than the number of classes defined in the first class definition.