[Question]: How to handle objects half out of frame when verifying in frigate+ ?

[JohannCR]

What is your question:

Sometimes, an object is detected while half out of frame. Is it ok to label it anyway or should we only use images where the objects are not cut off ?
For example a car sitting in the corner of a frame, visible at around 60 to 80% can be included and labeled ?
I know that if part of an object is hidden we need to tag its entirety, including the non visible parts to have a correct ratio, but this is not possible when the object is partly out of the frame.

[JohannCR]

Also the paid upgrade is "coming soon", does it mean it's not available yet, or maybe not in every country ?
That's too bad, I just spent a couple hours annotating a bunch of images thinking I would try it out 🤣

[hawkeye217]

Just email blake@frigate.video

[JohannCR]

Thanks, that settles this !
Only question left is how to handle cut objects. Any suggestion ?

[hawkeye217]

If a car is between 60-80% visible, yes, go ahead and label it.

If you haven't seen the docs on labeling and improving your model check them out here: https://docs.frigate.video/plus/improving_model

[JohannCR]

I did ;-) thanks for the confirmation on the cut objects.
Maybe we should add that to the documentation, people are bound to encounter some situations that raise questions (like if a snapshot labels an object in one place but the object has moved to a different place in the meantime, or when labeling faces is there a minimum size/pixel definition, do we label the whole head or just the face, stuff like that).
I know data quality is paramount for the training dataset (I work in the field^^), so since it's a shared dataset, documentation is important unless there are safeguards to detect wrong labels...

[NickM-27]

Not every image that a user submits is added to the base model that is used for other user's. Blake picks them by hand and validates / fixes those images. So submitting incorrect images would only affect a users own model. The process was outlined in detail at #9466

[JohannCR]

Wow that thread is very interesting, full of nuggets ! Thx
FYI made a pull request to update the docs