AI-assisted video anonymization, Human GT, and quality verification for real-world datasets, supported by customer-specific processing and review rules.
Visit Website · Watch Demo · View Technical Report
Send us a representative 1–5 minute sample from your own project.
We evaluate the footage under your project requirements and explain the expected processing scope, Human Verification requirements, deliverables, and pilot options before any larger commitment.
Origin Data Lab helps AI teams, dataset companies, mobility platforms, and CCTV solution providers prepare privacy-safe, quality-controlled video data for model development, evaluation, and commercial delivery. Our workflow combines:
- automated face detection and anonymization
- human verification for missed, uncertain, or high-risk cases
- Human GT, annotation correction, and dataset QA
- customer-specific masking, review, and acceptance rules
- delivery in video, image, YOLO, COCO, JSON, JSONL, or CSV formats
This repository presents a public V58 evaluation using dense, low-light urban traffic footage.
This public V58 evaluation shows automated face-anonymization output and is not a claim of perfect production coverage. Commercial projects can include an agreed Human GT or verification layer based on the customer's quality target, risk level, and acceptance criteria.
Our services are designed for teams that need privacy-safe, quality-controlled video data, including:
- ADAS and autonomous-driving teams
- AI dataset and annotation companies
- CCTV and smart-city solution providers
- robotics and mobility companies
- computer-vision research teams
- organizations preparing video data for model training, evaluation, or external delivery
Typical project requirements include:
- face and license-plate anonymization
- Human GT and missed-detection recovery
- customer-specific annotation or masking rules
- video and frame-level quality verification
- structured metadata and format conversion
- difficult footage involving low light, motion blur, occlusion, or dense scenes
|
|
|
|
▶ Watch the Public V58 Demonstration
The demo includes:- original footage
- automated anonymized output
- side-by-side comparison
- dense urban traffic
- low-light and partially occluded faces
| Metric | Result |
|---|---|
| Video duration | Approximately 1 min 29 sec |
| Processed frames | 2,661 |
| Detected face regions | 10,634 |
| Processing time | 238.3 sec |
| Processing speed | 11.169 FPS |
| Mean detection confidence | 0.637 |
| Evaluation environment | CPU processing |
These values describe this specific public evaluation and should not be interpreted as guaranteed performance for every dataset or environment.
| Large AI/BPO Vendors | Origin Data Lab |
|---|---|
| Standardized workflows | Project-specific workflows |
| AI-only processing | AI + Human Verification (where agreed) |
| Multiple communication layers | Direct communication with the project team |
| Large-scale engagement first | Representative sample evaluation first |
We do not assume that one public benchmark represents every camera, country, lighting condition, or operational environment. A private sample evaluation allows customers to assess the result using their own footage before considering a larger project.
Automation handles routine detections where conditions allow. Human verification can focus on missed faces, low-confidence regions, temporal inconsistencies, motion blur, occlusion, small targets, and other difficult cases according to the agreed project scope.
Each project can define its own:
- target objects
- masking strength and rendering method
- Human GT or human-review scope
- acceptance criteria
- output formats
- metadata requirements
- security and deletion conditions
Customers communicate directly with the team responsible for processing and quality review. This reduces handoff delays and allows requirements to be clarified before production scaling.
Human verification is applied according to the customer's quality target, security requirements, and agreed review scope.
Review tasks can include:
- missed-face or missed-object recovery
- low-confidence frame inspection
- blur and masking coverage checks
- temporal consistency checks
- bounding-box creation or correction
- annotation creation or correction
- customer-specific exception handling
- final delivery QA
Human-reviewed outputs can be separated from automated outputs and delivered with status information such as:
- automated
- human-reviewed
- human-corrected
- exception or unresolved case
Review capacity, access permissions, QA sampling, and delivery schedules are agreed according to project volume and security requirements.
A short private sample allows both sides to confirm technical feasibility, expected quality, security requirements, and delivery scope before a larger commitment.
Provide a representative sample together with:
- target objects to anonymize, annotate, or verify
- approximate total project volume
- required output format
- desired schedule
- Human GT or verification requirements
- security, retention, or deletion requirements
We review the footage conditions, target difficulty, processing requirements, and expected QA scope.
Before processing begins, we confirm:
- pilot deliverables
- processing and verification scope
- expected turnaround
- security and transfer method
- pricing or commercial conditions, where applicable
The sample is processed using the agreed automated workflow and, where included, Human GT, human verification, and quality review.
Depending on the agreed scope, delivery may include:
- processed video or image outputs
- QA findings
- detected limitations or edge cases
- metadata or annotation files
- HTML or PDF evaluation report
After reviewing the result, both sides can decide whether to proceed with a larger project and confirm volume, schedule, acceptance criteria, security controls, and pricing.
Request a Private Sample Evaluation
| Service | Scope | Example Outputs |
|---|---|---|
| Face anonymization | Automated privacy masking with optional human verification | MP4, MOV, image sequences |
| Human verification and GT | Missed-detection review, correction, annotation, QA | YOLO, COCO, JSON, JSONL, CSV |
| Dataset preparation | Metadata generation, restructuring, format conversion | Customer-defined schemas |
| Private pilot evaluation | Small-batch quality and workflow validation | Processed sample, HTML/PDF report |
Security and data-handling requirements are reviewed before each private pilot or commercial project.
The project agreement can define:
- approved data-transfer and delivery methods
- project-specific storage locations
- personnel authorized to access the data
- whether human verification is permitted
- retention and deletion schedules
- NDA and confidentiality requirements
- separation of customer project files
- restrictions on model training or secondary use
- customer-specific security documentation
Project-specific security procedures are defined before processing begins, including data handling, access permissions, retention, delivery, and deletion requirements where applicable.
The exact controls depend on the agreed project scope and available infrastructure. This public repository does not claim certifications or regulatory compliance that have not been formally verified.
Automated anonymization can be affected by:
- severe motion blur
- extreme low light
- heavy occlusion
- very small or distant faces
- reflections, displays, posters, or face-like objects
- rapid camera movement
- compression artifacts
- unusual camera angles
Automation alone may not fully resolve every situation. Human verification can be included according to the quality target and risk level agreed for the project.
Repository structure
privacy-processing-demo/
├── README.md
├── LICENSE
├── assets/
│ ├── logo/
│ │ └── origindatalab-logo.png
│ ├── charts/
│ │ ├── confidence_chart.png
│ │ └── detection_chart.png
│ ├── images/
│ │ ├── before_after.jpg
│ │ ├── before_after_003s.jpg
│ │ ├── before_after_008s.jpg
│ │ ├── before_after_030s.jpg
│ │ └── before_after_061s.jpg
│ └── thumbnails/
│ └── demo-thumbnail.jpg
├── demo/
│ └── origin_data_lab_face_anonymization_demo_v58.mp4
└── reports/
├── v58_public_evaluation.html
└── v58_technical_report.pdf
Origin Data Lab provides privacy processing, Human GT, human verification, dataset QA, metadata generation, and custom data-production support for AI and computer-vision projects.
We begin with a small, representative sample rather than asking customers to commit immediately to a large project.
The purpose of the pilot is to verify:
- performance on the customer's own data
- customer-specific processing rules
- Human GT or human-review requirements
- security and data-handling conditions
- expected deliverables, schedule, and cost
Production scaling is discussed only after the customer has reviewed the pilot output.
Email: contact@origindatalab.io
Website: https://origindatalab.io To request a private sample evaluation, please include:
- organization and project type
- data type and representative sample duration
- approximate total project volume
- objects to anonymize, annotate, or verify
- expected output format
- required Human GT or verification scope
- desired schedule
- security, access, retention, or deletion requirements
We will first review technical feasibility and project requirements. Pilot scope, deliverables, schedule, security conditions, and pricing will be confirmed before processing begins.
Origin Data Lab — Practical AI Processing, Human Verification, and Dataset Quality Support