[Frigate+] New labels toggled on for existing cameras and images maintain previous state

[pyrodex]

Love F+... Can't say enough and I love the fact new labels are getting added. I think its is a bit annoying though as we look at the life cycle of F+ and how many images folks are uploading that when we toggle on a new label in existing cameras with existing labels that EVERY older image now requires a review and looks like an OCD persons nightmare with that number now showing X images to be reviewed.

Could we make it so when a new label is toggled on for a camera that going FORWARD only new images will get that label for review and our pasts don't come to haunt us?

For example:

Someone merely toggled on a new label here and now they have a 2200+ image backlog to go back and review.

[heffneil]

I agree I can't add new labels without the fear of having to review thousands of images. I filter but those that are approved so having to reapprove all just isn't in the cards and messes up my workflow for future submissions.

[NickM-27]

Could we make it so when a new label is toggled on for a camera that going FORWARD only new images will get that label for review and our pasts don't come to haunt us?

You need to re-verify your previous images, if labels are missing then that will confuse the model during training when that label is not annotated. See this discussion for more info #18110 (reply in thread)

[heffneil]

ok sorry I guess we aren't understanding each other. I think its important to understand that by adding a label we arent ignoring the previously reviewed images for previous labels?

The second part - I don't review by labels I review but unreviewed images :) So its a ton of bricks.

[NickM-27]

I think its important to understand that by adding a label we arent ignoring the previously reviewed images for previous labels?

there is a button for this in the UI

The second part - I don't review by labels I review but unreviewed images :) So its a ton of bricks.

it is the exact same UI, you just select person in the dropdown...

[heffneil]

and then I have to pick car? and then license plate? And dog?

[NickM-27]

and then I have to pick car? and then license plate? And dog?

no, this is simply a filter for which images to show. It has nothing to do with what you annotate

[NickM-27]

I would kindly suggest actually trying what is suggested, instead of assuming that the recommended feature works in the most difficult to use way possible

[blakeblackshear]

Could we make it so when a new label is toggled on for a camera that going FORWARD only new images will get that label for review and our pasts don't come to haunt us?

You really can't. Imagine you are helping someone who lost their memory learn what dogs look like again.

First, you show them 100 images with a labeled dog in them. Then, you test their learning by showing them pictures without labels and ask them if they see a dog in the picture. When they guess some object is a dog incorrectly or miss a dog, they lose a point. When they correctly guess the location of the dog, they gain a point.

Now you want to teach them what a bird looks like too. You add another 100 images to the examples with labeled birds and dogs so you have 200 images total. What happens if you don't go back and ensure that the first 100 images also have birds labeled?

During testing, if the person sees a bird in the first 100 images, they will be penalized for an incorrect guess even though there was actually a bird present in the image. The answer key for that image didn't say there was a bird present. Now they are going to start getting confused. For some reason, they are losing points on their test when identifying birds in the first 100 images even if they look exactly the same as what they were shown in the second 100 images.

[heffneil]

I guess we are asking for the ability to mark them processed despite the fact that they truly aren't in their entirety. Thanks for the explanation. I dont' think we are trying to screw up the models just keep our processes of approving the same without a ton of already processed images fouling up our workflow.

[blakeblackshear]

All that would do is set the flag saying you reviewed the image for an object even though it's not true. Any future suggestions would be skipped for the object on that image because the data says that a person manually reviewed it already.

These older images will impact the models ability to recognize an object if they contain unlabeled objects. Your model may not ever be able to detect that object because of the missing labels.

The only thing I can think of is showing images that haven't been checked for any objects at all vs images that have been partially checked for only some labels.