The fine-grained classification work associated with her thesis isn't an _easy_ technical problem. The dataset, labeling, and architecture generated a few other first-author conference papers for her besides the thesis, so it's not like she just applied an off-the-shelf model to an existing dataset...