Labeling clarification


I realize that partial labeling is undesirable but I would like a little clarification. Consider that one is interested in two objects, say, dogs and cats. I can understand if the image has two cats and two dogs and only three of the four objects are labeled, this image would be classified as partially labeled. However if an image has two dogs and one cat and only the dogs are labeled, is this image classified as partially labeled? Having the cat unlabeled and part of the background should not be an issue for training dog objects from this image or does it? Does having an unlabeled object in one image corrupt the cat training with the other images? I guess I am inquiring how independent is the training from one image to another image. Does having an unlabeled object in a background of one image impact the training from labeled objects in other images?


Hi Skipper,

If you are training one model to detect both dogs and cats, then you want to label every instance of dogs and cats visible and recognizable to the human eye in all images in the dataset. From there, you would train one model.

If you are training one model to detect dogs and another model to detect cats, then you would only label dogs in one dataset/project, and cats in the other dataset/project. From there, you would train two separate models, or you can merge the datasets to train one combined model.