I thought it was kind of fascinating to see how people were reporting false positives on pets and vehicles occasionally – and then to stumble across a guy who mentioned that he was tagging EVERY pet and vehicle motion event as a “person,” because he didn’t want to risk missing anything (Which would obviously skew the effectiveness of the person detection.)
Specifically, that got me thinking about the idea of teachable AI. To a certain extent, I guess that’s what the person detection feature already is, except that I don’t think it’s using my personal data to teach itself, right? (Aside from the extent to which the videos I share through the beta app wind up getting factored into the global algorithm)
Anyway, I was just thinking about some of the amazing things we could do if we were able to teach our cameras custom AI. Obviously, this kind of feature would be advanced, and probably not for the faint of heart. I’m sure it would require the user to do a lot of tinkering in order to get things just right, and to build the dataset in the first place. But that doesn’t really scare me. I’m a web developer, so I’m sort of a hacker/tinkerer by nature. I think it could open the door for some pretty cool things – some features that are light years beyond what any other camera companies are providing.
For example, with custom AI specific to the camera, it might be possible to cut down on the false positives from headlights at night. I get why that’s a difficult problem, because each camera is different, and there are some headlights that ARE real motion (for example, If someone is coming up the driveway and the camera is blinded by the light.) But if it were possible to feed the camera real-world examples from the specific camera in question (i.e. “THIS is important motion, THIS is unimportant motion”) it would be possible to get more meaningful alerts. The same idea could be applied to trees swaying when there’s too much wind during the daytime – if it’s specific to the camera, and the camera’s position, it’s much easier for the camera to be smart about it.
It might also be possible to get information about specific people on camera (I know people have already brought up facial recognition in the forum ) or specific vehicles on camera. Obviously, one red Mustang looks pretty much like another, but if the camera knows that a red Mustang tends to be in front of the house regularly, it could send an alert like “Jane is home!” with reasonable accuracy , assuming someone else with a red Mustang didn’t park in your driveway. It could also learn to recognize the mail truck, UPS truck, FedEx truck, etc.
Anyway, I know some of this is probably shooting for the moon, but I love tinkering with stuff like that. If you guys are even remotely considering doing advanced stuff like that, I’d be thrilled!