Privacy Operations Open Source

MaskingOPS

End-to-end workflow for privacy-preserving video de-identification and quality assurance in behavioral research.

Mask it with MaskAnything. Verify it with MaskBench. Archive it FAIRly.

~86%
Pose preserved (blur)
7
Pose estimators
5
Masking strategies
FAIR
Data-ready outputs
The Challenge

Two problems, one workflow

Video data is central to behavioral research across psychology, communication science, and human-computer interaction. Yet privacy concerns create two interconnected operational challenges that limit data sharing and reproducibility.

Challenge 1: De-identification

How do you remove identifying information from video while preserving the behavioral signals your research depends on?

Challenge 2: Quality Assurance

How do you verify that your de-identification hasn't compromised downstream analyses like pose estimation?

The Pipeline

End-to-end in four steps

MaskingOPS integrates de-identification, quality assurance, and data archiving into a single operational workflow.

Step 01

De-identify with MaskAnything

MaskAnything

Upload your video, select a masking strategy (blur, pixelation, solid fill, contour, or skeleton), and let SAM2-based segmentation automatically detect and track people across frames. Use the human-in-the-loop editor to refine any tracking errors.

SAM2 backbone YOLO prompting Interactive refinement Docker deployment
Step 02

Benchmark with MaskBench

MaskBench

Run your raw and masked videos through multiple pose estimators. MaskBench computes standardized accuracy metrics (PCK, RMSE) and kinematic smoothness metrics (velocity, acceleration, jerk) to quantify the impact of masking on pose estimation quality.

7 estimators PCK & RMSE Kinematic metrics Automated reports
Step 03

Evaluate and Iterate

Review quality reports. If pose estimation accuracy falls below your threshold, adjust your masking strategy and repeat. The iterative loop ensures you find the optimal balance between privacy protection and research utility for your specific use case.

Quality threshold Strategy adjustment Decision support
Step 04

Archive FAIRly

Export masked videos with full provenance metadata: what was masked, which strategy, processing parameters, and quality scores. Deposit to DANS, institutional repositories, or other FAIR-compliant archives with confidence that your data is both privacy-safe and research-ready.

FAIR metadata GDPR compliant Provenance trail DANS / ODISSEI
Results

Masking strategy matters

Our benchmarking shows that the choice of masking strategy dramatically affects downstream analysis quality. Not all masking is equal.

~86%

Pose information preserved with Gaussian blur

~52%

Pose information preserved with pixelation

~31%

Pose information preserved with solid fill

Strategy Privacy Level Pose Quality Best For
Gaussian Blur Moderate
86%
Gesture & movement research
Contour Moderate-High
~70%
Spatial interaction studies
Pixelation High
~52%
Moderate privacy needs
Skeleton Overlay Very High
~45%
Maximum privacy with motion
Solid Fill Maximum
~31%
Clinical / highly sensitive
Components

Two tools, one workflow

MaskAnything

De-identification framework

Flexible video de-identification built on the Segment Anything Model 2 (SAM2). Supports multiple masking strategies, multi-person tracking, consent-aware processing, and human-in-the-loop refinement through an interactive web editor.

  • SAM2-based segmentation with YOLO pose prompting
  • 5 masking strategies: blur, pixelation, solid, contour, skeleton
  • Interactive editor for tracking error correction
  • Batch processing via CLI or web interface
  • Docker-based deployment for reproducibility
GitHub

MaskBench

Benchmarking framework

Open-source benchmarking framework for evaluating pose estimation quality on raw and masked videos. Integrates multiple pose estimators with standardized accuracy and kinematic metrics for systematic quality assurance.

  • 7 pose estimators: YOLOv11, MediaPipe, OpenPose, MaskAnything (4)
  • Accuracy metrics: PCK, RMSE per keypoint
  • Kinematic metrics: velocity, acceleration, jerk smoothness
  • Automated visualization and comparison reports
  • Extensible: add custom datasets, metrics, estimators
GitHub
Recommendations

Choose by research context

Different research scenarios call for different privacy-utility trade-offs. Here are our recommended configurations.

Gesture & Movement Research

Full-body kinematics must remain intact. Blur preserves the most pose information while providing moderate privacy.

Blur YOLOv11-Pose

Face-Focused Interaction

Facial features must be fully obscured while preserving body orientation and spatial relationships.

Solid fill Careful QA

Large-Scale Archiving

Processing entire corpora for long-term storage. Use default presets with batch processing for consistency.

Default preset Batch CLI

Clinical / Sensitive Data

Maximum privacy with human-in-the-loop verification. Every frame must be checked for residual identifiability.

Solid fill / Skeleton Manual QA
Compliance

FAIR data, GDPR ready

MaskingOPS outputs are designed for integration with research data management workflows. Every masked dataset includes provenance documentation, processing metadata, and quality assurance reports.

  • F
    Findable

    Persistent identifiers and rich metadata for masked datasets

  • A
    Accessible

    Compatible with DANS, ODISSEI, and institutional repositories

  • I
    Interoperable

    Standard output formats (JSON, CSV) and documented schemas

  • R
    Reusable

    Clear licenses, processing provenance, quality scores attached

GDPR Compliance

MaskingOPS helps researchers meet GDPR Article 89 requirements for processing personal data for scientific research purposes. The workflow produces documented evidence of data minimization and proportionate privacy measures.

Informed Consent

Masking is not a substitute for informed consent. MaskingOPS supports consent-aware processing where different participants can receive different masking levels based on their consent status.

Documentation

Learn MaskingOPS

Tutorial Paper

Step-by-step guide submitted to Behavior Research Methods, covering the full MaskingOPS pipeline with worked examples.

Under review — Behavior Research Methods

Video Walkthrough

Watch the complete workflow from video upload through quality verification to FAIR archiving.

Watch demo →

Masking Lab

Hands-on training events where researchers learn MaskingOPS with their own data and expert guidance.

View schedule →

Start masking your research videos

MaskingOPS is free, open source, and designed for researchers. Deploy with Docker, process your first video, and verify the results.