Learning without forgetting in Mask-RCNN

[SupreethN]

Is there a way to implement Learning without Forgetting in Detectron2?

More precisely, let's say I want to train a MaskRCNN to detect a new object, say balloon, using a custom dataset.
If I fine-tune the MaskRCNN weights originally trained on the COCO dataset, the Mask-RCNN can only detect this new object and forgets COCO objects.

I want to train the MaskRCNN using this custom data in such a way that it should detect the new object balloon and also humans. One straightforward way would be to use COCO images along with the custom dataset to retrain the entire Mask-RCNN for two objects, balloon and person.

Are there any other efficient or smart ways to achieve this task without retraining on the COCO dataset?