A YOLOv8x-Seg-based adversarial domain adaptation approach for construction and demolition waste instance segmentation

Abstract

Effective construction and demolition waste (CDW) sorting is essential for improving resource recovery and supporting circular construction practices. Automated sorting systems require robust visual perception across changing material appearances, lighting, backgrounds, and operational environments. While semantic segmentation supports material-level mapping, instance segmentation separates individual waste objects, making it more suitable for robotic sorting and object-level handling. However, instance segmentation models trained on one dataset often experience substantial performance degradation in unseen target domains due to domain shift. This study formulates DA-YOLOv8x-Seg, an adversarial unsupervised domain adaptation approach that extends a YOLOv8x-based segmentation framework by integrating gradient reversal layers and a unified domain classifier to learn domain-invariant features from labelled source and unlabelled target images. The experimental validation across two target domains demonstrates consistent improvements of approximately 10% in mAP50-95 under moderate and severe domain shifts compared with the source-only baseline. Benchmarking, ablation study, and feature alignment analysis further confirm improved robustness, stable parameter behaviour, and the advantage of the unified domain classifier over alternative configurations. Inference time analysis verifies that the adaptation branch introduces no runtime overhead because it is used only during training, thereby supporting more robust object-level perception for automated CDW sorting and circular construction waste management.