Panoptic segmentation

#2
by johnnv - opened
Image Processing and Computer Graphics Lab org

This dataset can also be used for panoptic segmentation, but the masks and also JSON are not provided in the mendelay data. I have made a tool to help with this conversion - CCAgT-utils.
I think the best way is to generate this data locally and upload it to this repo. How do you think it would be most appropriate to add panoptic segmentation option to the dataset here in the hub?

@lhoestq @mariosasko

There can be an additional configuration for this task :)

The dataset script can import from other python scripts if you want to include your converters' code in this repository:

from .converters import *

then maybe it's possible to do the conversion on-the-fly when generating the dataset in _generate_exmaples ?

Image Processing and Computer Graphics Lab org

Yes, it can be done this way, but this way it will be cached for users to just download it?

The converters code will be downloaded at the same time as the dataset script itself.

And the rest works as usual: once a user has downloaded and prepared a dataset with load_dataset, it will be cached locally

Image Processing and Computer Graphics Lab org

So, the conversion will occur locally, in run-time? This will be slow, due to the size of the dataset.

The conversion will run once when the dataset is downloaded and prepared. But if you reload the dataset it won't re-convert the data again, as it will use the cache.

If you think it takes too much time, maybe it would be worth hosting the converted data directly somewhere ?

Image Processing and Computer Graphics Lab org

I need to test how long this will take, try it on some collab machine to test. But I think this should be 5 to 10 minutes, which is fine.

Sounds good ! Let me know if you have questions or if I can help :)

johnnv changed discussion status to closed

Sign up or log in to comment