Recently, data augmentation techniques for training conv-nets emerge one after another, especially focusing on image classification. They’re always applied to object detection without further careful design. In this paper we propose COG, a general domain migration scheme for augmentation. Specifically, based on a particular augmentation, we first analyze its inherent inconsistency, and then adopt an adaptive strategy to rectify ground-truths of the augmented input images. Next, deep detection networks are trained on the rectified data to achieve better performance. Our extensive experiments show that our method COG’s performance is superior to its competitor on detection and instance segmentation tasks. In addition, the results manifest the robustness of COG when faced with hyper-parameter variations, etc.