My dataset has some images with whole-image classification annotations and some without (I don’t know why). How do you remove whole image classification annotations? Seems like you can only add them and once added you can only modify them. Am I missing something?
I’m really only using bounding box annotations. Whole image classification annotations are adding confusion into my dataset.
Can you provide screenshot? It shouldn’t have classification labels if the dataset type is not classification.
(new users can only put one embedded image in a post)
It’s actually fairly easy to replicate.
Here, in the classification mode, how do you remove this annotation? I can only choose to add or change the annotation
Why is the dataset type being switched after labeling?
(edit: my example image was a test with a brand new dataset, just for demonstration purpose)
Not sure why my dataset was originally switched to classification after detection labeling (I’m thinking bad data formatting while importing additional images), but is it relevant?
My point being whatever the reason why a whole-image classification label was added, it’s impossible to remove it and confuses the dataset from that point on.
In that example image, how can I remove the classification label?
@Chibbs if your dataset switched to classification by accident you can change it back to detect here, then the classification labels should disappear.
You’re also right we should be able to remove labels completely if it’s a classification dataset, we’ll work on adding that functionality.
@glenn-jocher what you’re saying is not what’s happening. switching back to detection does not clear the classification labels. I’ve just tried again using a brand new account and brand new dataset.
steps to reproduce:
here’s another example:. One label was created in classification first. Then switched to detection, added a label. Now there are two labels:
But only one detection label:
Got it thanks, those steps are pretty clear, I’ll try to reproduce again and work on a fix then.
Oh I see what you’re saying now! Classes added simply can’t be deleted (in one mode or after changing to another mode). Got it, understood. This is part of a wider feature we need: deleting classes.
I’ll work on a fix for this and let you know once this is done.
Yes, deleting classes (as a wider feature) would also work in my specific case, and would indeed be an interesting feature to have.
Deleting classes (as a wider feature) might not be enough for all cases where you need to remove specific classification labels from specific images, so that would also be a nice feature to have.
In the meantime, I understand that this is most likely not impacting the training of the model, but only confusing the user (me) when browsing through labels statistics for my dataset. Let me know if I’m wrong.
Thanks for taking the time to understand my issue, really appreciate it.
@Chibbs good news! We’ve shipped a new class delete feature as we’ve had a few users asking for it. You can see it in the Dataset Classes tab.
Let us know if this works for you!
