Retrying to exclude false positives

[grinco]

I'm getting high confidence false positives sometimes when detecting humans, it can either be caused by a shadow, a dog, or others. I am wondering if there is an easy way to filter those out. For example - have a configuration parameter indicating how many frames are sent to the API before the detection can be trusted. If my dog looks 80% like a human in one frame, but then it's missing in the second - probably it was a false positive. Can we implement something like that?

[LordNex]

I’m for this as well. I’m currently using a combo of frigate, DeepStack, CompreFace, and Double Take, and I still get false positives. With only 3 models trained DeepStack over CompreFace usually has the wrong entry. Thus the reason I added another detector to check against. That way both detectors have to come back positive before the rest of my automations move. On Jan 20, 2022, at 3:13 AM, grinco ***@***.***> wrote:  I'm getting high confidence false positives sometimes when detecting humans, it can either be caused by a shadow, a dog, or others. I am wondering if there is an easy way to filter those out. For example - have a configuration parameter indicating how many frames are sent to the API before the detection can be trusted. If my dog looks 80% like a human in one frame, but then it's missing in the second - probably it was a false positive. Can we implement something like that? — Reply to this email directly, view it on GitHub<#265>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ACGXNJA6JNGEEENBI3VQJBLUW7G2RANCNFSM5MMFR4FQ>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[robmarkcole]

In general, you need to raise the confidence threshold, say to 90%

Ensemble approach is interesting, am makes sense if some models perform better under different conditions, e.g. at night. However managing the ensemble can be complex