DualMatch: A Dual EMA Teacher for Underwater Semi-Supervised Pipeline Segmentation
Resumo
Underwater pipeline semantic segmentation remains a challenging task due to visual distortions, limited labeled data, and complex class boundaries. In this work, DualMatch is proposed, a semi-supervised segmentation framework based on a Dual Teacher architecture with Exponential Moving Average (EMA) models. Unlike prior approaches that rely on a single teacher, our method leverages two complementary teacher networks to reduce noise and enhance consistency in generating pseudo-labels. The powerful DINOv2-Small is adopted as the encoder, and our model is evaluated on a challenging underwater dataset. To support this research, a dataset comprising images from Remotely Operated Vehicle (ROV) and Autonomous Underwater Vehicle (AUV) inspections is proposed, with semantic labels manually annotated for five object classes. Experimental results show that DualMatch achieves the best performance in four out of five semantic classes, and notably reaches an IoU of 84.69% in the Pipeline class, outperforming all compared methods, including UniMatchv2 and Allspark. Overall, the proposed approach sets a new standard in underwater semantic segmentation by delivering consistent and high-accuracy predictions across complex classes, most notably for pipeline segmentation. The code and dataset are available at: https://github.com/EduardoLawson1/DualMatch
