I highly doubt AI is being used here to match up the Open Camera footage with the instagram pic. Identifing humans as an object within a frame is trivial. Identifying individuals requires a considerable amount of custom training which is more than what is available from an instagram pic and a few video frames. I'm not saying it can't be done, it's just not that easy. I'd bet $100 bucks that most, if not all, of these matches were accomplished using good old stalking (matching post times to manually reviewing video data). Would be happily proven wrong but without any source code provided I stand firm with my opinion.
Have a look at pimeyes.com. A single picture is enough, often even with sunglasses. Though that site requires input photos of slightly better quality than what you would be able to get from surveillance webcams. Then again, try it often enough, and use AI enhancement/upscale, and perhaps you can just use the site to replicate the project.
I suspect it's likely that there's a fair amount of effort involved, but I don't think you'd have to work too hard to substantially narrow your search with machine learning. Even just a small classifier that outputs clothing color and timestamps would make a remarkable difference. Identifying humans (which you can see in the videos) and checking how long they linger for could inform potential candidates for folks getting their photo taken.
Comments
I highly doubt AI is being used here to match up the Open Camera footage with the instagram pic. Identifing humans as an object within a frame is trivial. Identifying individuals requires a considerable amount of custom training which is more than what is available from an instagram pic and a few video frames. I'm not saying it can't be done, it's just not that easy. I'd bet $100 bucks that most, if not all, of these matches were accomplished using good old stalking (matching post times to manually reviewing video data). Would be happily proven wrong but without any source code provided I stand firm with my opinion.
Have a look at pimeyes.com. A single picture is enough, often even with sunglasses. Though that site requires input photos of slightly better quality than what you would be able to get from surveillance webcams. Then again, try it often enough, and use AI enhancement/upscale, and perhaps you can just use the site to replicate the project.
I suspect it's likely that there's a fair amount of effort involved, but I don't think you'd have to work too hard to substantially narrow your search with machine learning. Even just a small classifier that outputs clothing color and timestamps would make a remarkable difference. Identifying humans (which you can see in the videos) and checking how long they linger for could inform potential candidates for folks getting their photo taken.