| # ORION WWF1 — Unified Methodology (Anonymization + Labeling) |
|
|
| ## 1. Overview |
|
|
| The ORION dataset follows a **Dual-Engine Pipeline** to produce high-fidelity industrial data that is both private and intelligent: |
| 1. **Anonymization Engine**: Guarantees zero-leak privacy (Dual-pass AI + Human Audit). |
| 2. **Labeling Engine (Adjudicator)**: Produces multi-layer industrial annotations (HAR, PPE, Scene Graphs). |
|
|
| --- |
|
|
| ## 2. Anonymization Pipeline (Pass 1 & 2) |
|
|
| *Refer to the previous sections for the detailed 5-model cascade (YOLOv8, YOLOv11, InsightFace) used for the Dual-Pass Anonymization.* |
|
|
| ### Visual Destruction Standard |
| - **Blur**: Gaussian Blur (Kernel 99x99, Sigma 30). |
| - **Expansion**: 1.68x safety margin. |
| - **Verification**: Secondary scan for residual leaks with Vertical/Horizontal Flip TTA. |
|
|
| --- |
|
|
| ## 3. Industrial Labeling Pipeline (Orion Adjudicator) |
|
|
| Once the video is anonymized, it enters the **Adjudicator** pipeline for high-level semantic enrichment. |
|
|
| ### Layer 1 — Human Activity Recognition (HAR) |
| Activities are labeled using a temporal event model: |
| - **Scope**: 44 distinct activity segments across 10 clips. |
| - **Taxonomy**: walking, carrying_object, assembling, inspecting, cleaning, operating_machine, standing_idle. |
| - **Precision**: Sub-second start/end timestamps. |
| |
| ### Layer 2 — Safety Compliance (PPE) |
| A multi-category safety audit is performed for every agent: |
| - **Categories**: Head, Eyes, Hearing, Respiratory, Hands, Body, Legs, Feet. |
| - **States**: YES (Compliant), NO (Non-compliant), HIDDEN (Not visible from camera angle). |
| - **Validation**: 100% human-verified safety audit reports. |
| |
| ### Layer 3 — Physical AI Environment |
| Scene-level metadata is extracted for robotics and plant optimization: |
| - **Plant Context**: Floor type (concrete), Lighting (artificial), estimated area (m²). |
| - **Asset Hierarchy**: CNC machine state (Active/Idle), navigation hazards, camera mounting height/angle. |
| |
| ### Layer 4 — Object Detection (COCO Format) |
| To support ML training, keyframes are extracted and annotated: |
| - **Frequency**: 1 frame per second (310 total frames). |
| - **Annotations**: Bounding boxes in standard COCO JSON format. |
| - **Categories**: 1=Worker (Agent), 2=Industrial Machine (Asset). |
| - **Anonymization Consistency**: All frames are extracted from blurred video to ensure 100% privacy compliance. |
| |
| --- |
| |
| ## 4. Quality & Certification (Level 3) |
| |
| The **Level 3 Certification** implies: |
| 1. **AI Proposition**: All detections and labels are initially proposed by specialized AI models. |
| 2. **Human Adjudication**: Trained industrial analysts review and correct every single label. |
| 3. **Integrity Locking**: Final assets are hashed (SHA-256) to prevent tampering. |
| |
| --- |
| |
| ## 5. Unified Taxonomy |
| |
| The dataset uses a strict hierarchy defined in `orion_wwf1_taxonomy.json` and validated by the `orion_wwf1_fields_schema.json` to ensure consistency across the entire 622-clip mother dataset. |
|
|
| --- |
| *ORION – Industrial AI Data Lab | Pipeline Version: Orion Unified V5.2* |
|
|
