Unsupervised domain adaptive object detection based on harmonious learning

Abstract

Marine ports are critical hubs for global logistics, yet their complex operating environments and high-risk work characteristics pose major challenges for safety management. Traditional object detection methods rely on large amounts of labeled data, which are time-consuming and costly to obtain in port environments. To address this problem, we propose an unsupervised domain-adaptive detection framework, HLMix, based on harmonious model learning combined with a region-level sample mixing strategy. HLMix filters pseudo-labeled regions using a harmony measure and generates mixed samples by combining high-harmony regions from the target domain with source-domain samples for training. Furthermore, we design an unsupervised harmonious consistency loss and a harmony-weighting loss to enhance pseudo-label quality, fully exploiting the potential of hard samples in the target domain and improving the generalization ability of the detector. In addition, the information extraction capability of the detector is further improved by integrating the enhanced C3k2-CG module into the framework. Experimental results on the Hard Hat Workers → Port Night Operations and Sim10K → Cityscapes domain adaptation tasks demonstrate that HLMix significantly enhances detection performance, improves pseudo-label quality, and increases detection robustness. The proposed method provides effective technical support for safety detection in complex port environments.