Video data collection for AI, filmed to spec with per-participant consent

Avatar and lip-sync teams commission talking-head footage: framed faces, clean audio sync, and enough speakers, lighting setups, and cameras that a generated head does not inherit one studio's look. Audiovisual speech recognition teams want the same faces in worse conditions, because the visual stream earns its keep where the microphone struggles. Emotion teams commission facial expression data across elicited and spontaneous states, with the capture plan and label taxonomy designed together so the footage and its expression labels are specified as one deliverable.

Liveness and anti-spoofing buyers need genuine presentations recorded on the phone cameras their KYC flow will actually meet, alongside scripted attack material. Driver and occupant monitoring teams need in-cabin footage across seating positions, eyewear, and light that runs from noon glare to near darkness. The spec sheets differ, but a human video dataset stands or falls on the same operation underneath: recruit the right people, film to the sheet, check every clip.

A face on video is biometric-adjacent data: the footage identifies a person whether or not your model cares who they are, and under the GDPR biometric data used for identification sits in the special categories. Several widely used public face datasets have been withdrawn after scrutiny of where their images came from, because permission that was never collected cannot be added later. Whatever license text accompanies a scraped set, the people in the frames were not asked about model training.

Commissioning face data collection inverts that. Every participant signs likeness, biometric, and AI-training language that names the intended use before the camera rolls, and the signed record travels with the delivery. Spirelight runs the same recruitment, consent, and QA operation for video that it built for speech collection, so the rights review happens at scoping rather than after a due-diligence question.

What separates video data collection services is rarely the camera; it is what arrives at the end. Spirelight delivers a custom video dataset structured for a training pipeline: every clip keyed to a manifest entry with participant, session, device, lighting, pose, and approved-use fields, in a schema agreed before the first session. Annotation, from bounding boxes to expression labels, can be layered onto the same clips so footage and labels never drift apart.

Video datasets for machine learning fail on unglamorous defects, so QA checks for them per clip: face visibility and framing, focus and exposure, audio sync wherever the spec includes sound, and metadata completeness. Clips that fail are reshot under the same consent rather than patched in post. The check runs before delivery, so what arrives is the dataset rather than a snag list.