Built a synthetic data generation pipeline in NVIDIA Omniverse for computer vision using a custom-authored data center digital twin, semantic labeling, multi-camera capture, and controlled camera perturbation to generate annotated RGB, segmentation, and bounding-box datasets.

A modular data center + office environment was built using OpenUSD principles, including:

• Server racks and infrastructure zones
• Office workstations and monitors
• Support equipment (carts, cabinets)

Six strategically placed cameras capture distinct semantic zones:

• Server aisles
• Office workspace
• Support areas

Each camera generates independent datasets with controlled jitter to simulate real-world variation.

Objects were labeled using Omniverse Replicator:

• Servers
• Desks
• Monitors
• Chairs
• Equipment

This enables automated generation of training-ready annotations.

Each frame produces:

• RGB images
• Semantic segmentation
• 2D bounding boxes

Server Dataset View

Office Dataset View

Semantic Segmentation

• Multi-camera dataset generation
• Scoped camera jitter (controlled perturbation)
• Lighting variation support
• Structured dataset output for ML pipelines

• Replicator instability → resolved via staged execution strategy
• Camera targeting issues → solved with authored camera system
• Scene coverage gaps → resolved via multi-camera architecture
• Semantic labeling pipeline → implemented programmatically

• NVIDIA Omniverse USD Composer
• Omniverse Replicator
• OpenUSD (USD / USDA)
• Python

• Domain randomization (materials, lighting, layout)
• Large-scale dataset generation (1k–10k frames)
• Integration with training pipelines (PyTorch / TensorFlow)

This project demonstrates how digital twins can replace real-world data collection by generating scalable, labeled datasets for computer vision.