Modality Translation for Object Detection Adaptation Without Forgetting Prior Knowledge