Gautam Birat, Arashpour Mehrdad
Department of Civil Engineering, Monash University, Melbourne, Australia.
Waste Manag. 2025 Jun 1;200:114744. doi: 10.1016/j.wasman.2025.114744. Epub 2025 Mar 23.
Data annotation is a significant bottleneck in compiling instance segmentation datasets, particularly in the context of construction and demolition waste. Data augmentation has been shown to address this issue by increasing the diversity, instances, and complexities of data. While augmentation techniques for detection datasets are well-documented, a comprehensive evaluation of methods to improve the robustness of instance segmentation models is lacking. In this study, we developed and evaluated various data augmentation techniques on a publicly available dataset. Our findings indicate a 6% increase in mask prediction accuracy using the class balance method and a 4% improvement when combining real and synthetic data for training. Additionally, the mask prediction accuracy for minority classes increased by 30% using the augmentation techniques. This study demonstrates practical augmentation techniques to enhance instance segmentation performance with adaptation capability in any instance segmentation dataset within the context of waste management.
数据标注是编译实例分割数据集的一个重大瓶颈，特别是在建筑和拆除废物的背景下。数据增强已被证明可以通过增加数据的多样性、实例和复杂性来解决这个问题。虽然检测数据集的增强技术有详细记录，但缺乏对提高实例分割模型鲁棒性方法的全面评估。在本研究中，我们在一个公开可用的数据集上开发并评估了各种数据增强技术。我们的研究结果表明，使用类平衡方法时掩码预测准确率提高了6%，将真实数据和合成数据结合用于训练时提高了4%。此外，使用增强技术，少数类别的掩码预测准确率提高了30%。本研究展示了实用的增强技术，可在废物管理背景下的任何实例分割数据集中提高实例分割性能并具有适应能力。
