Skip manual image review in Frigate+ / Option to mark false positives for each annotated object

[d-dsn]

The reason we're going through every single image manually when submitting them to Frigate+ is to identify "False positives". We need to tell Frigate if a detected "person" is really a person, or if it's not.

This process could be significantly improved by this workflow:

1. Give the option for Frigate to automatically submit all the images to Frigate+ with automatic annotation, whether it's false or true positive
2. After the automatic annotation of all the found objects, user could then mark false positives as "red rectangles". We can now mark something as "difficult". We'd need to have new checkbox for "False positive".

I have my car correctly detected as a "car" with a correct rectangle around it. But when I submit the image to Frigate+, it quite often detects half of the car as "another" car. Now I have to manually delete this "other" false positive car from maybe 30% of the pics. There's no way to un-train the model for this false car.

[blakeblackshear]

When you choose "false positive" from within Frigate, it submits additional data used during training so that I can attempt to reproduce the false positive for improving the base model. There is no way for you the user to provide that additional data if you were to mark it a false positive in Frigate+ directly.

The partial cars with blue boxes you are seeing are the suggestions. These are generated by an entirely separate model that isn't relevant for the performance of your model. I have some improvements coming that should reduce the rate at which this happens.

[d-dsn]

Thanks for the clarification.
How about in the "Tracked Object Details", could we have all the recognized objects marked with rectangles, and the user could point out which are false positives? Would the "additional data" will be still available for sending to Frigate+?

[NickM-27]

The problem is Frigate does not keep the information for every detected object at every frame