Conversion to COCO format

#1
by kreouzisv - opened

Hi there,

I am working on a similar project and was wondering if you guys plan to provide the dataset in COCO format? Alternatively, would you mind sharing a bit more granular details as the file contents to assist in me converting the dataset myself?
image.png
This is a snapshot for a single image with the decompressed files, if you can provide with a short description of each and how to potentially link them, I would greatly appreciate the support.

Big Lab org

Hey @kreouzisv ,

we don't have the data in COCO format, but there is a conversion script in the accompanying repo that should include pretty much all the data you need to convert things into the format needed for object detection: https://github.com/js0nwu/webui/blob/main/scripts/screenrecognition/generate_dataset_web.py

I've already run that script on the entire dataset and saved the results here: https://drive.google.com/file/d/1FBbMaDmwipEXsKBODCaqrAn9X58Lwz83/view you should be able to select only the folders that are in the split that you want. There are also many other links to models and pre-processed data in this script here: https://github.com/js0nwu/webui/blob/main/downloads/downloader.py

Hope that is helpful and sorry I haven't had time to clean up the scripts and document them. But let me know if you have any other questions.

Big Lab org

Hey @kreouzisv

We've just uploaded a version of the dataset that has been converted to HF's parquet format with the pre-computed element locations and labels. https://huggingface.co/datasets/biglab/webui-7k-elements

Hey @Jsonwu ,

Thanks for the heads up and the effort. Already converted it myself so all good.

Big Lab org

Cool! Glad you got it figured out.

Jsonwu changed discussion status to closed

Sign up or log in to comment