WorldArchive/overlap-catalog-ego100k

Overlap catalog: Egocentric-100K

Perceptual fingerprints of builddotai/Egocentric-100K, computed with overlap at 4 fps on decoded pixels - so re-encoding, container swaps and metadata stripping do not hide a match.

Import this into a local overlap index and every dataset you are offered is screened against the public corpus:

overlap import block00 block01 ...     # or just the blocks you want
overlap compare vendor-offer.ovlm      # "81% of this is public Egocentric-100K"

What is here

• source revision: fae604b751b25337d6fd8c4c53e595910c28f68f
• all 29,966 archives hashed (2048 tasks, 25.4 TB of source video)
• 2,002,872 clips, 100,011 hours, ~1.3 billion fingerprints at 4 fps
• 16 blocks, each a complete, independently importable manifest with its own Merkle root - fetch only the ones you need
• archive-index.json maps every source archive to the block and part holding it
• provenance.jsonl maps clip to archive to sha256

Exact duplicates found in the source dataset

Fingerprinting the whole corpus surfaced something worth stating: Egocentric-100K contains 12,672 byte-identical duplicate clips (0.63% of the corpus), some appearing three times, and 44 archives that are exact copies of another archive. Those are stored once here, under the first path seen; the duplicate archive names are listed in archive-index.json under duplicate_archives with the archive they duplicate.

Honest limits

• No crop ladder (crop_hashes is null). Cropped or zoomed re-encodes of this footage will not be found by an index built only from this manifest - a manifest carries no pixels, so crop geometries cannot be derived from it.
• Exported at full density (--stride 1, 4 fps). A manifest strided below that loses recall badly against re-cut footage.
• algo_id=pdq2, prep_id=p1. A comparison against a different recipe refuses to run rather than returning a wrong answer.