Re:Inside Pornhub
by shakna on Thursday February 24, 2022 @01:40AM
(#30451361)
Most video-based perceptual hashes work by taking a series of thumbnails at set intervals, say every 5% of the video, assembling a collage, and then making the perceptual hash against the collage.
As you note, it isn't particularly reliable, but it does actually tend to work fairly well against the loss of video compression. However, it doesn't tend to work well when the video is recut, with time added or removed. There is some research trying to combine approaches for finding important keyframes (ML-powered thumbnail searching) with perceptual hashing [0], which is looking promising, but it's still fairly unreliable.
As you note, it isn't particularly reliable, but it does actually tend to work fairly well against the loss of video compression. However, it doesn't tend to work well when the video is recut, with time added or removed. There is some research trying to combine approaches for finding important keyframes (ML-powered thumbnail searching) with perceptual hashing [0], which is looking promising, but it's still fairly unreliable.
[0] https://doi.org/10.1007/s10207-013-0211-z