ML Retraining Control for Appearance-Change Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing unmanned driving technologies face challenges in maintaining accurate detection and control of moving objects when the appearance state of the target region, including pathways and surrounding areas, changes, such as due to soiling.
Innovation Solution
An information processing device that includes a change information acquisition unit to detect changes in the appearance state of the target region, and a training unit that updates a machine learning model using new training data incorporating the changed appearance state, thereby maintaining accurate control signal generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a machine learning model is trained with images before appearance state changes, then the model can operate with initial training data, but the detection accuracy decreases when the appearance state of the target region changes
Solution Approach 1:
The system performs preliminary detection of appearance state changes in the target region before they significantly impact detection accuracy. When changes are detected, the system proactively triggers retraining of the machine learning model with updated images reflecting the new appearance state, rather than waiting for accuracy degradation to occur
Solution Approach 2:
The system establishes a feedback loop where detection results are continuously monitored for signs of appearance state changes in the target region. When changes are detected through image comparison or analysis, the system feeds this information back to trigger model retraining, creating a closed-loop system that maintains detection accuracy through adaptive updates
2Reliability
If the machine learning model is continuously retrained with new images, then the detection accuracy is maintained, but the system complexity and processing time increase
Solution Approach 1:
The system detects appearance state changes preliminarily and triggers model retraining only when necessary, avoiding continuous retraining. This conditional approach maintains detection accuracy while minimizing unnecessary processing time and computational resources
Solution Approach 2:
The system monitors parameters related to the appearance state of the target region (such as image similarity metrics, feature extraction results, or detection confidence levels) and triggers retraining only when these parameters indicate significant changes, optimizing the balance between maintaining accuracy and reducing processing time
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing device includes a change information acquisition unit, and, when an appearance state is determined, using change information, to have changed, at least one of a training unit and a notification control unit. The change information acquisition unit acquires change information. The training unit trains a machine learning model using a plurality of sets of training data in which a training image including a target region after a change in the appearance state and a moving object is associated with a correct answer label. The notification control unit notifies an administrator of training request information for training the machine learning model.