So i have been submitting, share results, of events for quite some time now. Does anybody in wyze look at these as the events are ridiculous. Leaves rustling in wind triggers glass breakage, spiders trigger person events, shadows trigger person alerts, rain drops trigger person alerts, clouds trigger person alerts, you name it. This is beyond ridiculous, to the point I have had to disable all the features i paid for, cam plus, on the cameras. There is nithing AI about the software, its junk. Does anybody at wyze listen to concerns and review submitted events.
I don’t think anyone at Wyze reviews all the clip submissions. More likely there’s an automated process that sends the submissions direct to the AI trainer.
1 Like
Probably the whole company is replaced by A.I now.
1 Like
I can’t say what is done now, but I know they definitely had someone in the past review training material before it was included.
I also know that they do a new AI update monthly.
However, considering there are already hundreds of thousands to millions of submitted data already, I imagine that it can take a good amount of correction to shift the detection weights. I have seen progress over time as I’ve mentioned in other threads. I have had a camera detect my black cat as a person at night almost 100% of the time, and slowly reduce that frequency to where it’s rare to never now, but it took a lot of training submissions over several months and progress was slow.
Actually AI pet detection in my case is very impressive. And this isn’t my pet, it’s the neighbor’s cats, all 3 of them going through my yard now that I don’t have a dog.
1 Like
I have not seen any improvement in fact it is worse to the point i have had to deactivate all the ai detection which defeats the whole purpose of the cameras. Leaves rustling detected as glass breaking, really.
Mine is pretty good. I have 4 cats and it gets them really well. It only used to mess up on my black cat, but does pretty well with him now too.
It still sometimes tags my 2yo as a pet though, or sometimes some of my kids’ stuffed animals.
The only false person detections I really get now are around my pool with my V4 cams. I don’t know why the V4’s are having false person detections near my pool, but I suspect it may be related to the local/edge AI they have on the camera because none of the other cameras watching the same spot do it.
1 Like
I agree with @carverofchoice.
It’s been almost a year (?) since I started training the AI (machine learning). I think a common misconception is training Ai is fast. Just the opposite.
Due to potential unintended biases training AI cannot be totally automated so the once a month update sounds plausible to me. Someone/something has to review the entry and determine if it is true and UNBIASED.
As noted by @carverofchoice when I first turned on the feature it was terrible. It did not know what my dog was. I methodically submitted corrections. At least once a day, as much as 3 to 4 a day. It took about a month and I started to see results. My dog was identified as a dog and not a motion event. It has been very reliable.
I have also seen major improvemts in other areas including curtains waving, bugs flying by and extreme lighting changes.
It is my opinion that Wyze can preload hundreds if not thousands of images of what a pet is. Where they need help is what is NOT a pet. My dog in my environment may be missed/mislabeled until the Ai is trained. And there are millions of households like mine in the US alone.
Recently when clicking on Friendly faces I was asked to take a survey about friendly faces. It was a good survey and my answer to one question was if an average user would have the time to do what we are currently doing. I think not. There needs to be a relook at how Ai training is done. I do not know the answer but it is food for thought.
If you want to see results - - when you see an event mislabeled take a minute to correct the event and press the submit button.
2 Likes
I have been submitting corrections on a regular basis, but see absolutely no improvement. Its to the point of ridiculous. Middle of night leaves rustling and i get glass breaking, as well as rain drops and insects triggering person detected.
I’ve been thinking about your last post and I have a suggestion.
I have 10 cameras strategically placed on our property.
Driveway that is looking for motion, vehicles and persons.
Front yard. Looking for persons, packages, pets and dog barking
Doorbell. Looking for persons and packages
Inside front door. Looking for persons.
Living room. Looking for persons.
Dining room. Looking for persons.
Office. Looking for persons.
Garage. Looking for persons.
Backyard. Looking for persons, pets, dog barking.
If you notice, every camera is set to the bare minimum to satisfy my use case. I do not expect a vehicle in my backyard. Or packages inside my house.
If you haven’t already done so I would revisit every camera and only detect the bare minimum. Then see what happens. If camera x is detecting rain drops as humans then focus your feedback on that issue. See if you don’t see a positive improvement after 30 days or so. If you do then focus on your next issue.
As I said earlier, trading Ai is not fast–just the opposite.