FrameMatch — how it works (demo)

The real tool fingerprints video frames and compares them to find re-encoded, trimmed, or resized copies. It can't run OpenCV in a browser, so this is a faithful model of the two decisions that drive it: how often to sample frames, and how strict a match has to be. A synthetic "edited clip" is compared against its "source" below — move the sliders and watch precision and cost trade against each other.

Sample

N frames / sec

Fingerprint

pHash + ORB

Match

Hamming dist.

Score

frame hits ÷ total

Decide

confidence

Sampling rate — 2 frame(s)/sec

Lower sampling is far cheaper but can skip the frames that would have matched. The real tool defaults to one per second.

Match threshold — Hamming ≤ 10

How different two fingerprints can be and still count as the same frame. Too loose and unrelated frames match; too tight and a re-encode breaks the match.

Edited clip — 0 of 0 frames sampled

✓ = matched to source

Frames compared

Frames matched

Confidence

Relative cost

Model, not the engine: fingerprints here are stand-ins so the demo runs instantly and offline. The real pipeline computes perceptual hashes and ORB descriptors with OpenCV and matches them across CPU cores. The trade-off you're moving — recall versus cost, strictness versus robustness — is the real one.