“…To evaluate the detection performance on novel classes, we follow the same division on base and novel classes according to the existing work [19], i.e., 3 novel classes (airplane, baseball diamond, and tennis court) in NWPU VHR-10, and 5 novel classes (airplane, baseball field, tennis court, train station, and windmill) in DIOR. On the NWPU VHR-10 dataset, we base-train on all images that do not contain any object of novel classes, and then randomly select a very small train set containing both base and novel classes for finetuning, in which each class only has K-annotated images, where K equals 1, 2, 3, 5 and 10.…”