Oh, I can answer this to some degree. We had some Wyze employees confirm that TECHNICALLY almost every Wyze camera has some degree of local/edge AI on them now (I think it was like 2.5yrs ago…but might’ve been 1.5yrs ago). The problem is that SOME of them have a low confidence interval because they are limited on how much of the model they can fit on the camera. So the confidence interval isn’t good enough to allow local AI detections on many of the cams by themselves, HOWEVER, what the lower confidence interval model can still do is quickly identify more likely detections and make that a priority for the cloud AI to analyze first, thus making notifications much FASTER than just waiting for the cloud to analyze every frame in order and then report when it finds something.
So in this case, let’s take the V3Pro cam which has a full local edge model loaded on it and can function without the cloud aI. It has a good enough model loaded into it with enough resources that it’s confidence interval is high enough to do it all by itself. Then when you add cam plus to that, it will still do the person detection locally and notify you WAY FASTER about a person being detected (than most other camera models) , but then it sends that footage to the cloud for verification with an even higher confidence interval. If the cloud disagrees with the local model’s conclusion, the cloud version overrules it because it is more accurate. Thus, when you search events later, it wouldn’t include the false detection in those results because the cloud AI confirmed there was no person. Or the same with a false negative…the cloud AI may find that the smaller local/edge model missed reporting a person when it should have and now that event will be labeled person when the edge AI may have skipped it. Similarly, if the Edge AI isn’t quite sure yet, it can ask the cloud AI to make that analysis a priority and see if it is a person without waiting for the cloud to check everything else in sequential order first. Together, they are more accurate and faster than alone.
So, building local AI into the camera, even if it is locked behind a paywall is still beneficial because when the local and cloud models both work together, they are extremely effective and way faster overall. so there is still definitely a point, and technically every camera has a local AI model on it, but most of the older devices don’t have a full local model with a high enough confidence interval to even support working by themselves. But it’s my understanding based on the public messaging that at least the V3Pro, FLPro, and V4 DO have sufficient capabilities to do this, and if it does have a sufficient AI model on it, it should be allowed to function that way.
So the cloud AI doesn’t replace the local AI model, they work together to be way better.