I have a wyze cam outside, like I’m sure many of us here do. Person detection for me currently has almost no false positives anymore (shadows, etc., don’t trigger it as much as before). But when it’s dark (especially because of this time of year), many person events go undetected. So this is just a reminder to focus on submitting nighttime person events to Wyze for AI improvement. I think this is even more important to do because of the outdoor cam coming soon. So every day now I’m going through and submitting the ones where it made a mistake, but especially at night, since I really want to be able to find out if, say, someone broke into my car while I was sleeping.
It also currently has trouble detecting people further away.
Remember, it’s important to make sure they have a balanced dataset for training, like day/night, people in shorts / people in winter coats, besides people in more unusual positions, such as on a bike, in a wheelchair, etc, so don’t forget to keep submitting those videos.
I would love it if I could also help train the AI for vehicle, pet detection, etc., but I don’t know how I could do this currently.
TL;DR: Those of us with cameras outside need to keep submitting videos to Wyze to improve person detection, especially in unique situations, because the outdoor camera is coming out and the current algorithms are probably optimized for indoor use.
You may want to add a Sense PIR motion detector near your car to trigger the camera on any body heat approaching the car. Yes, the car itself will also trigger the sensor, but your chances of catching someone in the act at night are greatly enhanced. Also, only a 1-minute cool down for those.
You will need to otherwise blind that sensor to seeing the street, people passing on the sidewalk, neighbor’s pets & kids, etc. I do that with a overhang, plus turning the sensor away from my neighbor’s yard.
Some of the limitations you likely won’t be able to get around by submitting clips are:
The distance between the camera and the target area of interest where people will likely appear should be between 2 - 20 feet. Outside that range the clip will probably be rejected.
Avoid low-light conditions as it decreases the effective range of detection.
You can additionally buy an outdoor IR illuminator to brighten the outdoor area, which would help you with #2 above.
I’ve had person events missed where it was very obvious to me it was a person, it was just dark (as in grayscale, not exactly night) and they weren’t facing the camera. I think these at least should be detected. Also, with enough training data, it could learn to detect people at great distances and in low lighting.
Personal detections are 35% effective. It’s not that bad at false positives it that it doesn’t detect me as a person or the delivery man. Shadows do give me plenty of false motion alerts. Maybe with less Sun this winter my motion alert will be reduced. Luckily we can filter between the two. Too bad the PIR doesn’t have person detection.
Yeah, in terms of false positives, I think the algorithm is almost perfect. I’ve literally only gotten one false positive ever, from a weird-ish shadow. And during the day, I’d say there’s almost nothing missed for me, too. But I agree that at night, it sometimes misses people. I also try to submit it when I know it’s missed something.
I understand why it’s probably tougher for the algorithm to make out details in low light, and in black and white. That actually makes me curious to see what happens if I disable night vision. I wonder if accuracy would go up or down. All of my cameras point at areas that are illuminated by a porch light, all night. I currently use night vision, but I’ve disabled it before, and the video is definitely still usable. If it increases person detection accuracy, I may disable night vision altogether. I’ll give it a shot tonight and see what happens.
I do find it odd that if a person is more than 10 feet away, it detects 100% of the time, but if I am right in front of it, it does not. This is either day or night. Is this something to do with the detection settings?
Thanks.
You’re aware of the cooldown period, right? That wouldn’t make any difference with distance, but it definitely causes the camera not to detect “people” when you might assume it should.
If it is a distance issue, I wonder if it has to do with whether the whole body is viewable. For example, “person detection” isn’t exactly the same thing as “face detection,” although I would think it would have some overlap. But I think if you’re only seeing, like, part of a torso or a shoulder or something, it would be tough for the AI to determine that’s a human.
Hopefully, it gets improved. I’ve turned it off on my cameras because it never detects people. used to detect my cats and dogs as people but that ended with the last public firmware update. Now, nothing. Not even when I am deliberately going into the field of view. Maybe I’ll try again next year.
Check to make sure that it didn’t get turned off. Mine got turned off without my input. I think it happened with a firmware or app update, or when I enabled Complete Motion Capture. I’m not sure exactly when.
I used to get many false persons from shadows, coiled power cords, etc. Recently it’s more like person detect is off with only one detection from about 10 meters away. Each day many people walk by on the side walk across the street undetected. Even motion detection misses them with max sensitivity.
Several days ago on two V2 and seven Pan cameras at two different locations more than 10000 km apart… all had Show Wyze Logo turned on without my input.
Person detection isn’t as accurate if the cam can’t see the person. Now, under normal daytime lighting conditions, or within the range of an IR light, the person is properly illuminated. But at night without IR lights it isn’t expected to work well. Outside the range of IR lights would also be crippled by being more than 20 feet away. It can identify people under IR lighting.
You’re only getting person detection 35% of the time in daylight situations? Not related to cooldown, right? In daylight, my person detection is probably at least 95%. At night, it’s maybe 25-35%.
