Instance Segmentation of Exterior Insulation Finishing System using Synthetic Datasets

The quality inspection of adhesive of Exterior Insulation Finishing System (EIFS) is important because poor adhesive can lead to detachment of the insulation. Computer vision-based inspection stands out as a notable alternative. Recently, imaged-based deep learning model are widely used for the automated monitoring and inspection in construction field. To train the model, the relevant large datasets are essential. However, collecting datasets in the construction site is hazardous because of inherent risk of accidents. Also, synthetic datasets method which is one of alternatives to solve this problem are focused on fixed and regular shaped objects. To address these challenges, this study analyses the validity of synthetic datasets in terms of segmentation of adhesive in EIFS, which has irregular shape. For instance segmentation, the datasets were divided into two groups: (1) real datasets, composed of 100 actual photos, (2) mixed datasets, which combined 50 randomly sampled images from both synthetic datasets and real datasets. The mAP@50 of instance segmentation for real datasets and mixed datasets is 87% and 99%, respectively. This study prove that synthetic datasets can effectively train segmentation models, enabling the recognition of irregularly shaped objects and enhancing overall performance.