Consent and privacy when you record

Recording people at work, at home or in public captures more than the task. This lesson uses the privacy and ethics procedures of Ego4D, a public egocentric video dataset recorded by many universities and a company lab, to show what consent, review and redaction can look like in practice (Ego4D paper). Other projects set their own rules and the law differs by place, so this is a worked example, not legal advice and not any provider’s policy.

Two kinds of footage

Ego4D separates its video into two categories. The first is recorded indoors with informed consent collected from everyone in the scene, including their faces and voices. Only this kind is used for the benchmarks on audio-visual diarization and social interaction (Ego4D paper).

The second category, which makes up most of the video, did not have consent to capture identities, and includes footage recorded outdoors in public spaces. It has to be de-identified (Ego4D paper).

The rules each site followed

Each partner site developed its own policy, but they generally included four things (Ego4D paper):

• Follow the site’s own institutional research policy, including independent ethics review where relevant (Ego4D paper).
• Get informed consent from camera wearers, who can ask questions, withdraw at any time, and review and redact their own video (Ego4D paper).
• Respect the rights of others in private spaces and avoid capturing sensitive areas or activities (Ego4D paper).
• Follow de-identification requirements for personally identifiable information (Ego4D paper).

In short, video was captured either in a controlled environment with informed consent from everyone, or in public spaces where faces and other identifying information are blurred (Ego4D paper).

What participants were asked to do

At one site, participants were asked to choose a place at home with no identifying information, to avoid recording people who were not part of the study, and to avoid saying last names or other sensitive audio (Ego4D paper).

A few days later each participant was sent a copy of their video and asked to flag anything private, in picture or sound, that they wanted blurred or removed. The team edited it out, and also reviewed the footage for the faces of non-participants and for identifying details such as house numbers and license plates (Ego4D paper).

At another site, the team looked at every second of the video to exclude privacy-sensitive material such as license plates, phone screens and credit card numbers (Ego4D paper).

What redaction involves

De-identification used commercial software to detect faces and number plates automatically, and people reviewed every output for blurring that hit the wrong thing (false positives) and for faces or plates that were missed (false negatives) (Ego4D paper).

Other identifying content had to be found and blurred by hand. The authors list written names and addresses, phone screens and passwords, and tattoos (Ego4D paper). Accidental captures also count, such as an address on an envelope or the wearer’s face reflected in a mirror (Ego4D paper).

It is slow work. The authors report that videos recorded outdoors could take ten times the length of the video to redact carefully (Ego4D paper). In one sample pipeline, de-identifying 500 hours of video took 780 hours of manual labor (Ego4D paper).

Bystanders

One university’s ethics board allowed recording of incidental people in public settings, on the condition that collection was only of publicly observable behavior with no manipulation or direct interaction with them. The authors describe this as an assumed waiver of consent for those people (Ego4D paper). That is one board’s decision for one study. The general standard across the sites was that bystanders in public spaces are blurred (Ego4D paper).