AI Review question

photonclock
photonclock
in SecuritySpy

Playing with the AI Review feature for the first time.

I do not have Animal detection triggers enabled on any cameras, but I see the AI Review shows a remarkable collection of detected animals on my various exterior cameras. There's some good action shots in there of cottontails and road runners (yep, I live in the desert!) zooming across the field of view.

Questions:

1) I glean from this that the AI detection is fully independent of the selected triggers (all my cameras are set to (X) Motion (√) Human (√) Vehicle (X) Animal) so the AI model that's running is working to evaluate all object types? It's not like, three discreet models, for animals, or humans, or vehicles (ie, a performance benefit by evaluating only for one object type only?)

2) In AI review, there's an option to click on an image and choose "this is incorrect". What happens if I do that? Is the image transmitted somewhere, ie does it become model training data sent back to Ben Software?

3) When an object's detected type is "None" and it should be "Animal" (or Person/Vehicle as the case may be), would it be correct to flag it as "this is incorrect?" Does doing that that help the model accuracy on my machine, or otherwise act as training data for my local instance of SecSpy, or for Ben Software?

ScreenShot 2025-08-02 at 3.12.45 PM.jpg


Comments

  • Ben
    Ben

    To answer your questions:

    1. It's one model that outputs all predictions from one pass of the image through the neural network. So you get the animal predictions "for free" even if you aren't using them as triggers.
    2. The image is transmitted back to us (as long as you have agreed to share anonymous images in the General settings - if not you will get a pop-up message informing you that you need to enable this option to specifically agree to share images).
    3. Simply flag any such images as incorrect. They will be sent to us, and we will review and correctly label them, and use them to train the next iteration of the model that will be released in a future software update. Please do this when you can, as this really helps us to improve the detection accuracy of our model!

Categories