Learning goal
After watching, you understand what MaskAnyone does, where it sits in the SYNAPSIS pipeline, and can describe a masking job end-to-end. You don't need to have installed it yet — that's the deep-dive companion.
Transcript
Cold open
Stage one of the SYNAPSIS pipeline. This is where the masking actually happens. You start with sensitive video, you end with something you can share.
What MaskAnyone is, in plain terms
MaskAnyone is a free, open-source tool that takes a video of people and produces a de-identified version, while preserving the behavioural information you care about — gesture, posture, gaze, movement.
Under the hood, it's nine Docker services working together. A React web interface for researchers. A Python backend. Workers that run SAM2 for segmentation, OpenPose for pose estimation, and ANNY/Mesh3D for body modelling. You don't have to know any of that to use it — but it's good to know it's not a black box.
What goes in: a video with one or more people. What comes out: the same video, with the people masked in your chosen style, plus a pose-data file for analysis. Privacy preserved, science preserved.
Where it fits in the SYNAPSIS pipeline
MaskAnyone is stage one. It produces the masked outputs. From here, the masked videos flow downstream.
Stage two — MaskBench — measures whether your masked outputs still carry the information your analysis needs. Tutorial 03.
Stage three — MaskingOps — is the operations layer. Dashboards, batch processing, QA review across many videos. Tutorial 04.
Stage four — Archiving — packages the masked dataset with FAIR metadata and deposits it in DANS or a comparable repository. Tutorial 05.
A 90-second tour
The interface. You arrive at the dashboard. Upload area in the middle, project list below.
Drag a video onto the upload area. It appears as a project. Open it. Click on each person you want to mask. A coloured outline tells you the system has locked on.
Pick a masking style — blur, solid mask, silhouette, or skeleton-only. Tick extract pose data if you want a JSON file with skeleton coordinates for downstream analysis.
Hit start. The job runs in the background. With a GPU, expect 2 to 4 minutes per minute of video; on CPU, 10× slower.
When it finishes, scrub through the result. Look for slipped masks at fast movements, overlaps, and occlusions. Happy with it? Hit download. Three files come out: the masked video, the pose JSON if you asked for it, and a processing log — your reproducibility record. Keep the log. Always.
When you'd use which masking style
Quick rules of thumb for picking a style — your research question decides.
| Style | Privacy | Best for |
|---|---|---|
| Blur | Light | Studies that need facial expression or gaze. Use only if other identifying features are already controlled. |
| Solid mask | Moderate | The workhorse. Body and face fully covered with a solid block; movement and posture remain visible. |
| Silhouette | Strong | When you only need shape and trajectory. Almost no identifying surface left. |
| Skeleton only | Strong | Strips everything except the stick figure. Best when your analysis lives on pose coordinates alone. |
When to refine by hand
The automated pipeline gets you 80 to 95% of the way. The remaining gap — that's where you step in.
Three moves cover almost everything. Add a person the model missed — click them in their first visible frame. Remove a background figure caught by mistake — right-click. Adjust a sloppy boundary — drag the edge. Re-process is fast — only the affected frames re-run, not the whole video.
Recap
MaskAnyone in one breath. Drag a video in. Click each person. Pick a masking style based on your research question. Run. Review. Refine if needed. Download three files — masked video, pose data, processing log.
Two paths from here. If you want to install MaskAnyone right now, jump to the installation deep dive. If you want to keep the pipeline tour going, the next stage is MaskBench — proving the masked outputs are good enough to use.