Abstract:In the field of Object detection. It is Hoped that that the model trained in the domain with label-rich can be applied to other domain without label, but different domain distributions are always mismatch, Such a distribution mismatch will result in a sharp decline in performance.In this paper, for the purpose of improving the performance of object detection for domain tranfer, The Paper addresses the domain transfer on two levels:1) the globle-level transfer 2) the local-level transfer and apply different alignments to the aforementioned levels, specifically, globle-level adaptation with selective alignment and local-level adaptation with full alignment. The Paper construct the framework based on the benchmark SSD model, and devise two domain Adaptors on globle and local level respectively to alleviate the domain discrepancy. The Paper implement the two domain Adaptors in the form of domain classifiers through adversarial training manner on different levels and further apply a consistency regularization to the SSD mode and empirically verify the effectiveness of our method on various datasets,which outperforms the benchmark SSD and the other two domain transfer models by a large margin of above 5%-10% in terms of mean average precision (mAP) comprising both similar and unsimilar domain transfer scenarios.