Approach for Improving Segmentation Quality on Construction Sites Using SAM

Research on image analysis utilizing deep learning technology at construction sites is actively progressing. However, building the large datasets required to train deep learning models is a challenging task on construction sites. Therefore, this study suggests using the Segment Anything Model (SAM) to improve the segmentation quality of currently existing deep learning models without building additional training datasets. We chose Panoplane360 as a model that needed quality improvement and combined it with SAM. The experimental results showed an improvement in MIoU from 0.696 when using only the Panoplane to 0.714 when integrating the Panoplane and SAM. It means that it is possible to improve the results of deep learning models despite the data sparsity inherent in the construction domain. The following method can also be combined with other deep learning models, not just Panoplane, indicating its potential application in various fields.