[FR] Case for evaluating initial correctness of labelling

[senovr]

Proposal Summary

Abitilty to estimate correctness of ground_truth labels - with or without invoking a pre-trained models

Motivation

I have a question about following use case:
Imagine, that we have a dataset that have been already labeled by crowd (i.e., coco)
Apparently, there may be some mistakes ( wrong or missing labels) for different objects.
Do we have an option to evaluate initial correctness of labelling with fiftyone?
I was not able to locate such case in examples - the closest one is Digging into COCO

Example of the workflow:

• Extract patches from dataset
• Compute embeddings
• Compute "similarity" or uniqueness for each class of objects
• Return similarity.
• Most dis-similar labels can be filtered in app and evaluated visually
• Images with incorrect labelling sent back to crowd for re-labelling

Willingness to contribute

The FiftyOne Community encourages new feature contributions. Would you or
another member of your organization be willing to contribute an implementation
of this feature?

• Yes. I can contribute this feature independently.
• [x ] Yes. I would be willing to contribute this feature with guidance from
  the FiftyOne community.
• No. I cannot contribute this feature at this time.