Dual-Alignment Adaptive Teacher for Domain Adaptive Object Detection

There are significant differences between the labeled domain (source domain) and the unlabeled domain (target domain) in object detection. Domain adaptive object detection (DAOD) aims to adapt detection models to the target domain to improve detection performance in different environments. Existing methods have been proposed to address feature discrepancies between domains using the semi-supervised studentteacher framework, which have made effective progress. However, these methods suffer from classification bias and noise during feature alignment, resulting in their inability to effectively reduce feature discrepancies between domains. We propose DualAlignment Adaptive Teacher (DAAT) to solve the data bias problem in DAOD. We align the instance features of different classes in the source domain and the target domain with our Category-Embedding-Guided Cross-Domain Instance Alignment (CEGIA), enabling the model to effectively reduce the feature discrepancies between the source domain and the target domain. We introduce Covariance Alignment (CA) to uniformly identify diverse styles while preserving feature content information, thus further mitigating data bias. Experiments across multiple datasets and ablation studies show that our approach can effectively tackle data bias in DAOD and enhance the model's detection performance.