I’m creating a single-label classification dataset and in the dataset, I have a lot of pictures which I want unlabelled. When I upload the images by default these images are classified as Unlabelled but I cannot add them to the dataset with the default Unlabelled label (which I want to) unless I change the label of the image to anything else and then back to Unlabelled. This adds a lot of extra time to the annotation.
Could this be changed so that I can add Unlabelled images to the dataset without doing the extra step of swapping the label to something else and then back to the Unlabelled (default) label?
It’s not possible to add unannotated images to a dataset to prevent unlabeled images from tainting a training dataset.
Single-label classification models are designed to always produce one label, regardless of the image. However, what you can do in instances where you don’t want anything labeled is to create a “null class” or a class that’s basically equivalent to a null example in object detection models.
For example, a classification model I’ve made in the past identified people existing in an image by having two classes: people and no_people. You could do something similar by creating a “else” or “null” class.