Deep learning models tends to rely on training datasets, so when new data are entered, performance degradation occurs. Domain adaptation (DA) is a learning technique that reduces the domain shift between a labeled training dataset (the “source dataset”) and an unlabeled target dataset to prevent performance against the target dataset. Although great effort has been achieved in single-source domain adaptation for object detection (DAOD), multi-source domain adaptation for object detection (MSDAOD) has not been studied much yet due to the challenge of considering Domain Shift between Source and Target domains as well as Domain Shift between Source datasets. Recently, previous studies designed Discriminator as many as the number of source datasets to solve each domain shift and conducted alignment for all instances. However, separate alignment makes latent spaces as many as source datasets. So, it is difficult to align all the datasets at once. This motivates us to propose a novel method for multi-source detector adaptation based on multiple binary cross entropy. We build a single latent space that can align the target domain and all the source domains at once. Major cause of performance degradation in DAOD is alignment between data with large domain shift. So, we design the alignment in inverse proportion to domain shift. We experimentally verify the effect of our method on two evaluation protocols.
