Yeah, as another said, currently any motion anywhere in the entire frame triggers the AI to scan the ENTIRE stationary frame to see if any pattern it has been trained to recognize is present anywhere in the frame. In my understanding it’s not scanning to look at only the things that are slightly different between 2 or more separate frames, but anything in any frame that resembles the objects it’s been trained to look for.
The motion tags themselves are totally separate. Something can be a motion event and the AI can then separately find a vehicle within one of the frames of that video. The green line highlights are tagging areas where there was any pixelation change from one frame to the next, and it will only follow those pixel changes (color or greyscale differences between 2 frames), and that tracks “motion” in this way, but the AI likely just analyzes a single frame, and finds a car and tags it. It could be trained to only tag “Car” if 2 or more frames are found to have that “Car” on slightly different pixel locations (movement), or whether the “car” is in the exact same pixels between 2 frames (stationary). It should be relatively easy to train the AI to recognize that difference simply by comparing 2 or more frames and only reporting on it if there is a pixel change.
I think this feedback could be very useful for Wyze AI team to consider adding an option such as an option to “only detect moving objects, ignore all stationary objects” which people could then select if they don’t want tags for stationary cars, ignore statues, etc. It would also ignore any person sitting very still through an event (ie: my cat is laying down on the porch, I walk outside, but my cat doesn’t move, so it only detects a person (me who is moving) but not “Pet” (because my cat held still and didn’t change enough pixels)…though whenever my cat does move it will tag pet again. So while it might not show me every video my cat is in, it would then still show me the start and stop of whenever my cat enters and exits the frame.
I think that would be great feedback and training they can give the AI, to only add a tag when the identification for the object can be seen on different pixel locations between multiple frames (real movement). I hope they consider this. It should solve the issue of shadows causing vehicle detection labels because even if a shadow causes movement (pixel color changes between frames), the vehicle would still be in the same pixel locations, so there would be no movement and no reason to inform of a vehicle. I like the idea.