I work at a company that focuses on automating the data labelling process for computer vision. It is clear that generating massive amounts of labels, either by hand or automatically, without the ability to ensure a consistent level of quality across the dataset, is a problem. Which is why we are investing in automating the QA process for training data so mistakes like these don't happen:
That's interesting, I took the opposite approach when realizing how bad label quality can be and built a site where everything is manually checked by moderators.
https://blog.encord.com/post/automating-the-assessment-of-tr...