AI Region-aware pre-training for open-vocabulary object detection with vision transformers – Google Research Blog