Does AI ever 100% recognize?

[SOB]

My 10 year old daughter asked what I thought was an interesting question about our cameras and AI recognition/detection. She asked if our cams [running CPAI 2.5 and looking only for ‘person’ and ‘vehicle’] ever would recognize at 100%? After explaining to her why AI would likely NEVER be 100% I still thought it was interesting enough to get thoughts here. The highest I get is about 94% for both person or vehicle, sometimes. Usually more around ~70% depending on the conditions. Thoughts?

What’s your highest recognition percentage you’ve gotten?

 

[mat200]

My 10 year old daughter asked what I thought was an interesting question about our cameras and AI recognition/detection. She asked if our cams [running CPAI 2.5 and looking only for ‘person’ and ‘vehicle’] ever would recognize at 100%? After explaining to her why AI would likely NEVER be 100% I still thought it was interesting enough to get thoughts here. The highest I get is about 94% for both person or vehicle, sometimes. Usually more around ~70% depending on the conditions. Thoughts?

What’s your highest recognition percentage you’ve gotten?

In theory AI should be able to get very close to 100% unless it is a purposeful ruse ( example manikin ) .. otherwise autonomous driving cars will not ever make it.

 

[tigerwillow1]

I've been griping about this like crazy on the CPAI thread . People and vehicles seem to get the best recognition, sometimes too good. I've had trees and shrubs identified as people, and a propane tank and greenhouse identified as vehicles. Animals are the worst, with the wrong ID happening more often than the correct one, and a lot of complete misses. My two favorites are a herd of deer identified as one deer, plus pigs, cows, and a bicycle. Plus a bird in flight identified as a dog. I'm expecting it to get better with time. It's never going to be 100%. Sometimes I have a small or distant animal trip the IVS and I have to spend a minute or so finding it, and I'd never find it with just a static image, which is what the AI uses.

 

[wittaj]

If you mean for an individual trigger, yes I have got a few 100% (15 just today) and lots of 99%.

If you mean will it ever be 100% accurate all the time, I think it can always be fooled.

While using IVS rules don't give a %, I have a few cameras that have never false triggered. And over the course of a year, all of my cameras using IVS combined will be less than a handful of false triggers.

 

[tigerwillow1]

And over the course of a year, all of my cameras using IVS combined will be less than a handful of false triggers.

And for the opposition report, when snowing over the past few days, I had maybe a hundred IVS triggers per day, with CPAI thinking a tree or snow covered bush was a person in about half of them . Good thing it's quick to plow through the clips with BI.

 

[wittaj]

And for the opposition report, when snowing over the past few days, I had maybe a hundred IVS triggers per day, with CPAI thinking a tree or snow covered bush was a person in about half of them . Good thing it's quick to plow through the clips with BI.

That is wild - knock on wood or maybe I am lucky, but mine have never triggered during snow or rain unless there was a person or vehicle in the view:

 

[SOB]

If you mean for an individual trigger, yes I have got a few 100% (15 just today) and lots of 99%.

If you mean will it ever be 100% accurate all the time, I think it can always be fooled.

While using IVS rules don't give a %, I have a few cameras that have never false triggered. And over the course of a year, all of my cameras using IVS combined will be less than a handful of false triggers.

I was referring to the individual trigger being 100%. I have stood directly in front of my camera, on a nice clear day, with a nice clear picture, and AI is still only ~90% sure. Are you using Deepstack or CPAI? I'm curious, how you have your AI alerts setup - mine sends 3 images 2 seconds apart to be analyzed. Not sure if that helps or hurts the accuracy? Heck, should I really care as long as I'm not getting false triggers?

And for the opposition report, when snowing over the past few days, I had maybe a hundred IVS triggers per day, with CPAI thinking a tree or snow covered bush was a person in about half of them . Good thing it's quick to plow through the clips with BI.

I also get a ton of triggers in the rain/snow but without AI identifying anything (person/vehicle) they just get ignored. Never had a snow-bush person (yet)!

 

[wittaj]

I am using CodeProject and depending on the camera field of view and day or night, it will send anywhere from 2 to 15 images per event. I also have DoNotCancel being used so that it picks the best/highest percentage image for the alert thumbnail.

But I am also only analyzing for people or vehicle or plates and not all the other potential objects it can do AI on.