I had a few hand gesture datasets with multiple hand classes, all annotated. I needed to clone some images from them into a new dataset, but I noticed in one of them, ~100 of the right class annotated images did not appear. I therefore cloned the right images from another one of my datasets, which had the correct ~200 amount. I cloned all the images of forward, left, and right into a dataset called nano hand gestures. I generated an augmented dataset and noticed that many right class images lost their annotation and also noticed that about 100 of the right images did not show up in the Dataset tab, showing that this bug happened again. I replicated this bug again when I created nano hand gestures 2 to try to stop the 100 images disappearing:
I believe that this may be the reason why a model I trained on one of the original hand gesture datasets had really bad accuracy on the right class. Why does this happen? I would provide more images but can’t because I’m a new user.
Sorry to hear you’re having issues. As I understand, you’re having issues with annotations not showing up. Cloning images on Universe has two options: Importing the images & annotations, or just the images.
Can you confirm that when you cloned the images, you selected the option to clone both the image and annotations?
Yes, I used Import Images and Annotations. I tried it again on another set nano hand gestures - cloning for bug test by cloning the annotated right hands from old hand gestures 3, but many just lost their annotations the same way as previously.
I think the “hiding” of the right hand images from not showing as right nor showing as Marked Null has to do with the loss of annotations, but it’s weird because when they lose their annotations, they should be marked as Marked Null.