How Reverse Video Search Works: Embeddings, Time, and Copy Detection
Summary
Google reverse-searches photos and stopped there; a copied video is never the same file (crops, speed changes, re-encodes, sticker overlays, screenshots of screenshots). Reverse video search solves it in three moves: embed every second of video (Meta's Video Similarity Challenge caps descriptors at 512 dimensions per second, matched by inner product), align the descriptor sequences in time to localize the shared segment with timecodes, and train descriptors on copies (SSCD-style self-supervision) so edits land next to the original in embedding space. On Meta's VSC23 benchmark: 60.5 uAP for a strong image descriptor alone, 87.2 detection and 91.5 localization uAP for the winning video systems. Mixpeek runs reverse video search in production over customer libraries. Explore: mixpeek.com/reverse-video-search
About this video
Google reverse-searches photos and stopped there; a copied video is never the same file (crops, speed changes, re-encodes, sticker overlays, screenshots of screenshots). Reverse video search solves it in three moves: embed every second of video (Meta's Video Similarity Challenge caps descriptors at 512 dimensions per second, matched by inner product), align the descriptor sequences in time to localize the shared segment with timecodes, and train descriptors on copies (SSCD-style self-supervision) so edits land next to the original in embedding space. On Meta's VSC23 benchmark: 60.5 uAP for a strong image descriptor alone, 87.2 detection and 91.5 localization uAP for the winning video systems. Mixpeek runs reverse video search in production over customer libraries. Explore: mixpeek.com/reverse-video-search