I am trying to implement object detection on satellite images. I have a annotated dataset, but the images are large and the model accepts only 416 x 416 size inputs. How can I pass small parts of the image in the network, ensuring that the annotations are retained. Also, how to merge these results at test time?