Dataset Sample Hub
Given an arbitrary archive of footage from a multi-camera shoot, we discovered which files belong to the same real-world event and aligned them on a common timeline. An agent-driven discovery pass proposes candidate groupings from filenames, folder structure, and embedded metadata (EXIF, professional timecode); audio fingerprinting then confirms and precisely time-aligns the members.
Sample
Fields on the left describe every attribute; the example on the right is real JSON from one asset in this dataset.
Each pair-clip is one matched span across two different camera recordings of the same real-world event. source_1 and source_2 point to the two source assets and their time ranges; framewise_sim reports the cosine-similarity band (mean / min / max) across the matched frames.