How Hamsa recognises a repost that has been cropped, mirrored or covered in captions, and why it needs three methods.
A repost is rarely an exact copy. It has been re-encoded by a platform, cropped to a phone screen, mirrored to dodge detection, covered with captions, or screenshotted. Comparing files byte by byte finds nothing. So Hamsa compares what the frames look like, three different ways.
One keyframe, three fingerprints. Each catches what the others miss, and none needs the video itself.
A perceptual hash shrinks a frame to 64 light-and-dark bits. Re-encoding and resizing barely change it. It is fast and exact, but a heavy edit breaks it.
An image embedding from an open model (SigLIP) turns a frame into 768 numbers that describe the scene. It still matches under captions and overlays, but on its own it would also match a different video of a similar river.
Keypoints (ORB) are distinctive corners and edges. When dozens of them line up in a consistent pattern between two frames, they are the same shot, not just a similar one. This is what stops the second method from fooling itself.
Mirrored frames are checked too. When there is a match, the checker sees the post's frame and the original side by side, with the original's date and source. They do not have to trust Hamsa; they can see it.