UI RefExp task

[ivelin]

Congrats on the great work with DocQuery, Impira team and congrats on joining Figma.

Have you looked at transfer learning DocVQA into UI RefExp? Natural language Referring Expression for the location of the bounding box of a UI element.

I have had some success with Donut, but it is converging slower than I anticipated. After a couple of weeks on a single A100GPU, it is still improving slowly with IoU at 47%. About 10% improvement in the last week. IoU (intersection over union) of predicted vs ground truth bounding boxes seems to be a reasonable validation metric for this task.

Curious what your experience with the task might have been and whether I am approaching the problem from the wrong angle.

Here is my Huggingface space with links to fine tuning notebook.
https://huggingface.co/spaces/ivelin/ui-refexp

Regards!

[ankrgyl]

We have not! Let us know if you try it out though. Curious to hear how it goes.