Learning goal
After watching, you can upload a video, select the people to de-identify, choose an appropriate masking style, run the job, monitor it to completion, and download the result.
Transcript
Cold open
From raw clip to a de-identified, downloadable file. Six steps. Let's do it together.
Step 1 — Upload
Open the MaskAnyone interface. Click New Project, or just drag a video file onto the upload area. We're using the sample gesture-study clip — about 30 seconds, one person, 720p. Plenty for a first job. Once the upload finishes, your video shows up in the project list with a thumbnail. Supported formats: MP4, AVI, MOV.
Step 2 — Select people
Click the video to open the editor. The first frame loads. Click on each person you want to mask. A green dot marks where you clicked. A coloured outline shows the model's first guess at the segmentation. Scrub through to a few keyframes to make sure the tracking stays on the right person. Tracking failures show up as the outline jumping to the wrong body — flag any frames where that happens for later refinement.
For videos with multiple people, select them all in the first frame. Each gets a different colour so the system can keep track of who's who.
Step 3 — Pick masking style
On the right, you'll see the settings panel. The big choice is your masking style. Four options, ordered from least to most aggressive.
Blur keeps facial expressions roughly visible — good for emotion or gaze work. Solid mask is the workhorse for general behavioural research. Silhouette keeps the shape and movement but loses everything else. Skeleton only strips the person to a stick figure — maximum privacy, only kinematics survive.
Tick Extract pose data if you want a JSON file with skeleton coordinates for analysis — that's the file you'll feed into your statistics pipeline.
Leave output format on MP4 unless you have a specific reason not to — it's what every player and editor speaks.
Step 4 — Run and monitor
Click Start Processing. The card flips to show a progress bar and an estimated time remaining. Speed depends on hardware. With a GPU you'll see 5 to 10 frames per second — a one-minute video processes in 2 to 4 minutes. CPU-only is roughly ten times slower.
You can queue more videos and walk away. A batch of overnight processing is normal — set it up, leave it, come back to results in the morning.
Step 5 — Review the result
When the project flips to Complete, click it. You're in the result preview. Scrub the whole thing — don't just spot-check. Three places to look hard:
Fast movement. That's where tracking tends to drop. If you see the mask jump off the body, mark it for refinement.
When people overlap. Make sure the colours don't swap — that's identity confusion, and it's the most common subtle failure mode.
Occlusions. When someone walks behind a chair, behind another person, behind anything — does the mask resume cleanly when they reappear?
Happy with it? Hit Download. You get three files: the masked MP4, the pose JSON if you ticked that box, and a processing log with the exact parameters you used. Keep the log — that's your reproducibility record.
Step 6 — Refine if needed
If something needs fixing, click Edit. The refinement tools let you patch problems without rerunning the whole video.
Three common moves: add a person who appears mid-video — click them in their first frame. Remove a background figure who got detected by mistake — right-click. Adjust a sloppy boundary — drag the mask edge with your mouse.
Hit Re-process. Only the affected frames re-run — it's fast, even on CPU. Then re-review.
Recap
Six steps. Upload. Click each person. Pick a masking style and tick pose if you want it. Run. Review — fast movement, overlaps, occlusions. Refine if needed.
You've made a masked video. Next up: how to prove it's actually de-identified — visually and quantitatively. That's the Quality-check deep-dive, or stay on the pipeline tour with MaskBench.