Applying the knowledge of an object detector trained on a specific domain
directly onto a new domain is risky, as the gap between two domains can
severely degrade model's performance. Furthermore, since different instances
commonly embody distinct modal information in object detection