SCAN: Cross Domain Object Detection with Semantic Conditioned Adaptation

The domain gap severely limits the transferability and scalability of object detectors trained in a specific when applied to novel one. Most existing works bridge by minimizing discrepancy category space aligning category-agnostic global features. Though great success, these methods model with prototypes within batch, yielding biased estimation domain-level distribution. Besides, alignment leads disagreement class-specific distributions two domains, further causing inevitable classification errors. To overcome challenges, we propose Semantic Conditioned AdaptatioN (SCAN) framework such that well-modeled unbiased semantics can support semantic conditioned adaptation for precise adaptive detection. Specifically, crossing different images source are graphically aggregated as input learn an paradigm incrementally. is then sent lightweight manifestation module obtain conditional kernels serve role extracting from target better adaptation. Subsequently, integrated into well-designed Conditional Kernel guided Alignment (CKA) module. Meanwhile, rich knowledge transferred Graph-based Transfer (GST) mechanism, category-based feature space. Comprehensive experiments conducted on three benchmarks demonstrate SCAN outperforms large margin.