DSCA: A Dual Semantic Correlation Alignment Method for domain adaptation object detection

Yinsai Guo, Hang Yu, Shaorong Xie, Liyan Ma, Xinzhi Cao, Xiangfeng Luo

Abstract: Highlights•We propose a Dual Semantic Correlation Alignment (DSCA) method for domain adaptation object detection. The DSCA takes full advantage of context and class correlation semantics information to align object semantic information in source and target domains.•The Context Correlation Semantic Alignment (COCSA) including the Local Context Correlation Semantic (LCCS) and Global Context Correlation Semantic (GCCS) modules is proposed to align the context semantic information at the image level.•The Class Correlation Semantic Alignment (CLCSA) comprising the Class Semantic Decoupling (CSD) and Dynamic Adaptation Graph Convolutional Network (DA-GCN) modules is proposed to align class semantic information at the instance level.•We conduct extensive experiments showing that the proposed DSCA comprising the COCSA and CLCSA modules in DAOD outperforms the state-of-the-art on four challenging benchmarks.