Advanced data augmentation techniques to enhance instance segmentation dataset for construction and demolition waste...

Area of Science:

• Computer Vision
• Machine Learning
• Waste Management

Background:

• Data annotation is a major challenge in creating instance segmentation datasets, especially for construction and demolition waste.
• Data augmentation can enhance dataset diversity and complexity, but its application for instance segmentation robustness is underexplored.

Purpose of the Study:

• To develop and evaluate data augmentation techniques for improving instance segmentation models.
• To address the bottleneck in data annotation for waste management datasets.

Main Methods:

• Evaluation of various data augmentation techniques on a public instance segmentation dataset.
• Application of class balance and combined real/synthetic data training strategies.

Main Results:

• A 6% increase in mask prediction accuracy was achieved using the class balance method.
• Combining real and synthetic data improved mask prediction accuracy by 4%.
• Minority class mask prediction accuracy saw a substantial 30% increase with augmentation.

Conclusions:

• Data augmentation techniques effectively enhance instance segmentation performance in waste management contexts.
• The proposed methods offer adaptability for various instance segmentation datasets in waste management.
• Addressing data scarcity through augmentation is crucial for robust waste management AI models.