The problem I am having is that to my understanding we need to annotate all objects of all classes on the images we want to train (or fine tune) our YOLO on. This is because YOLO compares labeled classes against other parts of the image, so if the other part of the image contains class we need to detect, this will confuse the model, right?
My question: What if one class we need to detect has very few appearances and the only appearances of this class are on the images where other classes are visible too? Because this means that if we add more data with the examples of undersampled class we are automatically increasing imbalance because on these pictures there are objects of the abundant classes that we also have to label.
Is it possible to maybe manually cover (in proverbial "paint") appearances of the classes we have a lot of samples of? Maybe YOLO doesn't have big problem with unbalance?
