WCO fw v4.17.4.427 and VDB (can’t make it tell me fw version), iOS app v2.46.1(4)
I have shared and shared and shared videos from several cameras via the “Help us improve AI by sharing this one video with Wyze” over the last year, yet my cameras continue to fail to recognize the most obvious vehicles and people, though occasionally they succeed. The video doorbell sometimes doesn’t recognize that it’s a person who rings it!
Yes, Event Recording/Smart Detection/Person Detection and Vehicle Detection as well as “Detect Wyze AI Events” are turned on, and for the WCO, the detection area encompasses the whole image of the car or person.
But the AI is not improving! Does sharing videos help anything, or is it a waste of time?
Wyze updates their AI IQ once a month. Because the source material database is so large, it can take months for a lot of shared videos in particular environments to make a difference for that particular individual’s use case. Say that there are 100,000,000 images in the database, and I submit 100. My hundred will certainly help, but they can’t drive an immediate drastic change against all the other data. They’re only 0.000001% of the input. It will be more effective if lots of other people also submit similar situations where they are experiencing something similar, and then the AI will give more weight to those particular conditions. I had 2 particular cams that kept detecting my black cat as a person at night nearly 100% of the time. and it took several months of constantly submitting videos of this before that finally got fixed down to less than 100% of the time, and then after long enough, it dropped down to nearly 0% false detections. It sure felt like it took forever though.
Also, keep in mind that the cameras are optimized for detections within about 30 feet. They can sometimes detect objects farther away, but the confidence interval sharply drops after that, particularly as there are fewer pixels involved.
Many other things make a difference, including lighting conditions, glare/reflection, light/shadow contrast, and many other things. This may be why it seems inconsistent where sometimes it detects something and the next time it doesn’t.
It is frustrating that it takes so long, but I do know that Wyze definitely uses the videos and they do new updates on the AI every month.
I’ve been relabeling all my skunk “Person” videos for a few months and finally Wyze started recognizing them as Pet. So yeah in this one case I think it made a difference.
I on the other hand have the opposite experience: three or four months ago it seems to have been more consistently accurate; in the last month, it seems more consistently inaccurate.