[Support]: Would uploading a certain number to plus(and het the model) of false positives allow to remove a object mask

[ronbruins]

Describe the problem you are having

I have a camera on our front door pointed at a place where we put out bikes. Frigate keeps detecting a bike as a person, whenever there is (other) motion, I already have an object mask that should prevent this, but as the bike moves when it is used, I have to modify the mask and the area also is meant to actually detect persons, so I don't want to have a too large mask.

Would uploading a certain amount of false positives to my frigate+ and then created a new model, prevent this bike from being detected as a person?

Version

latest

Frigate config file

NA

Relevant log output

NA

FFprobe output from your camera

NA

Frigate stats

NA

Operating system

Other Linux

Install method

Docker Compose

Coral version

USB

Network connection

Wired

Camera make and model

reolink

Any other information that may be helpful

No response

[nwish]

I’d say there’s a good chance. A lot of people are reporting success in that regard with Frigate+, myself included. You’d essentially be teaching your model the the bikes are not people.

[ronbruins]

So in theory, anything that gives a constant amount of false positives, when used to upload to F+ it would allow to get rid of object masks. Any idea how many images would train the model accordingly?

[blakeblackshear]

It's going to vary significantly from user to user. True positives near the bike will also help. I would suggest reading this if you haven't already: https://docs.frigate.video/plus/improving_model

[ronbruins]

Yeah read that and that's what I have been doing mainly for our cats who now after using a trained model are not identifying as dogs anymore (I hope they are ok with that ;-) ) I didn't had true positives as the bike is pretty much in the front, so difficult to get true positives. I just kicked off another model request, so let's see what this one does and I'll take it from there. Thanks for the answers.

[nwish]

You could also do these things:

• trigger true positives near the bikes by walking around there and submit then to Frigate+ along with your false positives
• use min/max area and/or min/max ratios if applicable - for this to work, the bounding boxes for the bike(s) will need to be different in pixel size and ratio than the person boxes