It might be possible to reuse training proposals, e.g. in cases where the user wants to find and annotate one class of interesting objects after another. Offer a button in the "annotation candidates" stage to go back to the "training proposals" stage. When this is done, all annotation candidates are deleted, all training proposals are unselected and the image patches of the training proposals are generated again. The user then can select a different set of training proposals and restart the instance segmentation.
It might be possible to reuse training proposals, e.g. in cases where the user wants to find and annotate one class of interesting objects after another. Offer a button in the "annotation candidates" stage to go back to the "training proposals" stage. When this is done, all annotation candidates are deleted, all training proposals are unselected and the image patches of the training proposals are generated again. The user then can select a different set of training proposals and restart the instance segmentation.