Ten to fifteen years ago, computer vision was already fingerprinting video with landmark detection (SIFT, region segmentation, Canny edges, biomimetic color cor

Ten to fifteen years ago, computer vision was already fingerprinting video with landmark detection (SIFT, region segmentation, Canny edges, biomimetic color correction---or simply converting to grayscale). It helps that you'll almost always have a unique audio stream that closely tracks the video stream which can also be fingerprinted and time-scaled and/or time-shifted. It's likely they're already doing machine learning. Why wouldn't they? With what essentially boils down to content recommendation as a way to increase ad viewership targeting an audience likely to have short site visits, wildly different tastes and habits, and a constant influx of new users and new videos, they probably already fingerprint videos (and users) as a method to immediately link new users and new videos to an existing pattern maximizing ad revenue. Another questionable method which might cut down on abusive/unlawful behavior would be shadowbanning offending uploaders once you have a reliable fingerprinting method. Upload a video that's close but not exactly matching a fingerprinted video? Thanks, here's the URL. Privately, we're not adding it to the index for anyone outside your subnet until someone can manually verify the two streams are unique. It's probably even possible to spoof the view count and other properties based on similarly fingerprinted clips, so only the people who make a living from aggressively reuploading are aware of the shadowban. As a bonus, if such a fingerprinting/shadowbanning combo cuts down massively on casual/bot/opportunistic reuploading, the people who continue to modify the video _and_ audio stream enough to bypass the filter could be (just a theory) more likely to be malicious/abusive, and these are the people you definitely want to exclude from your community