Is there a way to prevent a cam v3 from detecting a waving flag as a car or person?
Keep submitting the videos as false detections and submit what should’ve been detected in that video. They do monthly updates to training the AI, and eventually it helps.
Also, are you absolutely sure it is the flag being detected as those things? It could be that the waving flag triggers an event, but that something else inside the frame has a general shape that could be getting detecting incorrectly, something (including a shadow or pattern on the ground) with a round shape that appears to be on top of something else (person), or something box-shaped…I guess the flag is sometimes box shaped, so that might be the car issue, and the flag pole might have a ball shape at the top of it? That would make sense.
Sometimes shifting your camera position or angle can make a difference, detection zones blocking out the flag area can help too. I would definitely recommend submitting the events for a few months and eventually see improvement. I had something similar at my last residence, and over time it got better and better until it was basically never happening anymore. The AI training did work over time.
A motion detection box and message positively identified the flag as a car. Repeatedly. There were no other moving objects in the field of vision at anytime during the flag-triggered events. There were actual cars in the frame but they were parked and standing still at all times. There is no doubt the AI was in error.
I will keep submitting the videos for AI improvement as you suggested.
Doesn’t matter if the person is dead or the vehicle has been parked for 5 years, if in view of the camera you will get person/vehicle tagged in the events video if something else sets the cam off. I turned my vehicle notifications off because every time the wind blows and a tree in view of the cam sways I get vehicle alerts because my truck is in the cam frame.
As Antonius mentioned, the AI isn’t technically detecting moving objects. My understanding is that what happens is:
- locally on the camera, if it notices a significant number of pixel changes between the still frame pictures of picture 1 and picture 2 (Frame 1 and frame 2 in the video), then it registers that as movement…that includes lighting/shadow changes or real movement, or any number of things.
- The camera then uploads those pictures (making a video) to cloud storage.
- The AI server grabs the pictures from the cloud storage to analyze them and see if it recognizes any objects in the picture…
- The Camera continues to check if there are pixel changes between pictures/frames and if there are it keeps uploading pictures/frames until they stop having differences between the pictures.
- The AI does NOT really pay attention to whether anything is moving. It’s basically analyzing a still picture. If it sees a parked car in that picture it will tell you there is a car in the video even if it isn’t moving. That may explain why you are getting vehicle notifications. If there is a tree branch shadow that doesn’t move at all but kind of looks like a very loose shape of a human, it will tell you there is a human even though it wasn’t moving at all. The AI is not analyzing what is moving at all, only the camera really analyzes whether there is movement, then the AI analyzes the uploaded frames and sees if anything in the picture at all resembles anything you asked it to look for (not just the thing that was moving). Therefore, if there are parked cars anywhere in view, the AI will almost always tell you it detected a car even though it never moved.
Wyze is working on updating this though. They told us in some public discussions that they are working on having the AI learn to ignore parked cars and only report cars if they are moving (ie: if their position changes between 1 picture and another picture that it analyzes). This will be wonderful. This is the main reason I turn off notifications for vehicles though, because there are always cars somewhere in view so even if motion was caused by a leaf blowing in the wind, it still told me there was a car, so I stopped requesting those notifications.
I hope that helps make some sense. The motion is only to determine when to start or stop the recording of a video and has nothing to do with the actual detection from the AI, since it doesn’t take motion into account at all in the early versions we have yet.
Still, Wyze said that submitting videos like you’re going to do will help them have sources to train the AI with, so doing that will be really helpful.
So if I understand what you’re saying, in fact, car, people and pet motion detection is not actually car, people and pet motion detection. The AI is comparing frame one to frame two and if there’s a certain threshold of difference of pixels between the two, it defines that as motion and triggers an event regardless of whether there’s a car, person or animal in the frame? Oye. If so, that’s not very useful.
But thanks for the thoroughness of your reply. It was very helpful.
You genuinely cracked me up. Thanks!
Yes, that is a good summary of what I was explaining.
This is the way almost all camera companies’ AI detection currently works (not just Wyze).
Thus, there are some people who have a statue in their yard that may get detected as a person in almost every video even though the movement was something else entirely.
That’s why it is called person detection or vehicle detection and not actually person motion detection or vehicle motion detection, because it doesn’t matter if the object itself is the thing in motion as long as any kind of motion anywhere triggers the even initially…but it does work if it was the person or vehicle that was in motion. It’s just not a requirement…yet…
Wyze said they have been working on training the AI to function differently though, such as training it to only care about cars that are moving, and limit the AI to scanning something that is specifically within a programmed detection “Zone” instead of anywhere in view. Those are things being worked on, but aren’t fully implemented yet.
Anyway, you clearly understand how it works pretty well now, and that was my objective.
My first day into my Wyze Cam journey and have already learned a lot thanks to you and Antonius. Thanks!
(I’ve been experimenting with zone detection. I’m hoping that bears fruit.)
That will be a great option to block out the flag area so that it’s not triggering an event so often. You’ll still record that area when something happens elsewhere in the detection zone, but at least the flag will stop initiating an event as often.