Run inference on own picture with externally inputted object labels and bounding boxes

[aliencaocao]

I have some images that contain a mixture of seen and unseen object classes. I have my own custom object detection model based on YOLOv5, and it is able to output bounding box and class label and confidence. Is it possible to input these YOLOv5 results into Oscar+ and thus only use the text generation part of Oscar+ to generate a caption of the image? Original image with bounding box can be inputted, but I do not want Oscar+ to do the object detection part as my own model take care of some unseen objects.