Exploring Using Segment Anything Model to Assist Ground Truth Preparation for Timber Scaffolding Segmentation

Scaffolds are temporary structures, made from steel or timber, to assist the construction at a high level. Although scaffolds play a vital role in the progress and safety of the project, less consideration from site practitioners is paid to them. Moreover, various researchers have usually tackled and focused on difficulties for steel scaffolds, even though timber scaffolds were also implemented in numerous regions where the material is available. Therefore, there is a need for supporting tools for timber scaffold supervision, which image segmentation from deep learning may be able to fill this research gap. Nevertheless, preparing deep learning requires a gigantic amount of training data with ground truth that is very labor-intensive and time-consuming to create. Currently, the Segment Anything Model (SAM), which has already been trained with millions of images, can highlight numerous objects without further training and might be able to assist this troublesome work process, but this possibility has not yet been explored. In this study, thirty sample images were duplicated and formed into two datasets for timber scaffold segmentation, one dataset was labeled the ground truth manually and another dataset was prepared with the assistance of SAM. The result showed that SAM can assist this novel work procedure, reducing 43% of the ground truth preparation time with a 0.828 F1 score. This research result can encourage other practitioners to utilize SAM as a supporting tool for creating the ground truth of timber scaffold segmentation.