Generalization of Construction Object Segmentation Models using Self-Supervised Learning

To evaluate the safety of construction site workers, deep learning models recognizing workers and safety equipment in construction site images are widely used. However, it is frequently observed that deep learning models based on supervised learning methods do not work well for unseen data in other domains having different visual characteristics. To address this issue, a novel method for generalizing semantic segmentation models was proposed. This method adopts two strategies: a domain adaptation method based on self-supervised learning and a copy-paste data augmentation. Source domain data with annotations (workers and hardhats) and target domain data without annotations are used for model training in a self-supervised learning scheme. The proposed model showed an improved generalization capability in semantic segmentation without annotation data of the target domain.