Show HN: UniversalDataTool – open-source collaborative data labeling

[seveibar]

https://github.com/UniversalDataTool/universal-data-tool

[a_c]

May I suggest to have some sample datasets in the demo. Am a bit confused as to what the tool can do as I don't mingle data frequently

[seveibar → a_c]

Absolutely, will add some larger/easier to use datasets.

There is a button to import cat images in "Samples -> Import -> Import Cat Images" which you may find interesting for testing Computer Vision stuff

[rg2004]

to make this more effective, add keyboard shortcuts

[seveibar → rg2004]

for sure- a couple seconds saved per sample will amount to many hours of saved time even for a small dataset.

I created an issue to track interest in keyboard shortcuts, but I've heard this feedback from the friends I had beta-test so I think it's gotta be next on the agenda :)

https://github.com/UniversalDataTool/universal-data-tool/iss...

[djohnston]

very cool, i was wondering what open source data annotation tooling existed, and coincidentally this popped up at the same time.

[erezsh]

Does it have an API for adding labels programatically?

I couldn't find any documentation, or even installation instructions..

[mkl]

> Scales to tens of thousands of data points per dataset

Can you clarify what this means? That sounds pretty small.

[seveibar → mkl]

You can load tens of thousands of images into the web app at a time. Unfortunately after 100,000 I start to see performance issues- especially with collaboration.

I'd like to fix this in the future- but with a web application these limits are hard to break through.

[mkl → seveibar]

Yes, that should definitely be fixable. I have a simple Flask image labeling app (whole images, not items within them like this), and it has no upper limit. Millions is just as quick and easy as dozens. It's so primitive I don't even use a DB, just one flat file per 10,000 images, that can then easily be processed with grep etc.

[kwerk]

How does this compare with Label Studio or Doccano?