Deep networks have shown remarkable results in the task of object detection.\nHowever, their performance suffers critical drops when they are subsequently\ntrained on novel classes without any sample from the base classes originally\nused to train the model. This phenomenon is known as catastrophic forgetting.\nRecently, several incremental learning methods are proposed to mitigate\ncatastrophic forgetting for object detection. Despite the effectiveness, these\nmethods require co-occurrence of the unlabeled base classes in the training\ndata of the novel classes. This requirement is impractical in many real-world\nsettings since the base classes do not necessarily co-occur with the novel\nclasses. In view of this limitation, we consider a more practical setting of\ncomplete absence of co-occurrence of the base and novel classes for the object\ndetection task. We propose the use of unlabeled in-the-wild data to bridge the\nnon co-occurrence caused by the missing base classes during the training of\nadditional novel classes. To this end, we introduce a blind sampling strategy\nbased on the responses of the base-class model and pre-trained novel-class\nmodel to select a smaller relevant dataset from the large in-the-wild dataset\nfor incremental learning. We then design a dual-teacher distillation framework\nto transfer the knowledge distilled from the base- and novel-class teacher\nmodels to the student model using the sampled in-the-wild data. Experimental\nresults on the PASCAL VOC and MS COCO datasets show that our proposed method\nsignificantly outperforms other state-of-the-art class-incremental object\ndetection methods when there is no co-occurrence between the base and novel\nclasses during training.\n