Versatile Teacher: A class-aware teacher-student framework for cross-domain adaptation

Runou Yang, Tian Tian, Jinwen Tian

Abstract: Highlights•We propose a novel teacher–student framework that not only excels at bridging the domain gap but also stands as the first to integrate vanilla instance-level alignment with a one-stage detector, demonstrating a promising practical outlook.•We devise the Class-aware Adaptively Pseudo-label Selection (CAPS) mechanism, aimed at enhancing the reliability of pseudo labels by applying class-aware treatment.•We leverage pseudo labels as saliency matrices to guide targeted instance-level align- ment for one-stage detectors, expanding the applicability of teacher–student framework in domain adaptation.