Fully automatic annotation

Hi again :slight_smile:

if I don’t have time to manually or semi-automatically label all images is there a way to label the remaining images automatically with the AI assistants (so not looking at each image and accepting the assistants proposal)?

Cheers,

Lucca

Hi Lucca,

I’m glad that you’re using Hasty so much :wink: So I assume that it’s working out well for you?

Yes, this is possible, but you can only unlock the functionality when our Assistants converge to a loss below 0.5. We do this to protect users from destroying their data-set on accident.

You can check if you unlocked automated labeling on your project if you go to project settingsautomated labeling. To get there you have to keep using the Assistants.

We’re working on a view so that you can see how far the models are, at the moment. If you want to know it already know, though, drop me a PM and we’ll look it up in the backend.

Cheers,
Tobias

Hey Tobias,

yes, Hasty proves to be useful tool :slight_smile:

It would be nice, if the auto-labelling was possible at any point. I see two use cases for it. The one I mentioned above. I want to use the imperfect labels because I can’t get to a loss below 0.5. The second use case would be I just want to have a prediction on every image before working on it. That saves a lot of time. I can just set it as ‘Done’ when the prediction is to my liking or go straight into editing. I don’t have to run the assistant first and accept the prediction.

Why would the dataset be destroyed if the images automatically labelled were labelled as ‘Auto-labelled’?

Cheers,

Lucca

Hi Lucca,

regarding to your question how it’d destroy a dataset – let’s say you labeled 1k images already. You could export those anytime and train a model on it. Now you use automated labeling on 4k more images in the dataset and it produces bad results. Now, if you do an export of the data now, you’ll add noise to your model training and you’d need to spend time cleaning the 4k images first.

Of course, they’re workarounds for this. And we have tons of ideas from a UX perspective how to solve this, that you don’t run into this risk and don’t need to find some workaround. We just haven’t built it yet, because we’re focusing on some major things to come out on the ML side, first. I hope you can understand this.

Coming to your second idea of automatically creating the suggestions. It’s actually not too bad, we’ll discuss it internally. Did you discover that you can use keyboard shortcuts as well to speed up the process? So, for example, if you’re using the Instance Segmentation Assistant you can generate the predictions by just pressing I and then enter to accept all of them. I hope this helps a bit.

:v: Tobias