Synthetic Data Generation

• Built procedural scene generation workflows in Blender to create large-scale synthetic datasets
• Rendered paired outputs:
  • RGB images
  • ID-color (viewport) masks for surface segmentation
• Enabled controlled variation in environments (lighting, materials, composition)

Custom Annotation Tool (Gradio-based)

• Developed an interactive annotation tool for converting synthetic renders into YOLO instance segmentation datasets
• Implemented dual-group system:
  • Wall (foreground structures)
  • Background (environment surfaces)
• Designed patch-based and global color selection workflows using ID masks

Advanced Annotation Features

• Connected-component selection for accurate surface-level segmentation
• Morphological mask expansion for better coverage
• Non-destructive editing via rectangle eraser
• Undo system supporting iterative annotation workflows

Dataset Export Pipeline

• Exported annotations in YOLO instance segmentation format (polygon-based)
• Implemented contour extraction using OpenCV (hole-aware polygons)
• Normalized coordinates and standardized dataset output
• Built verification layer for previewing annotations before export

Scalability & Workflow Efficiency

• Designed batch processing workflow for large datasets
• Automated pairing of render + mask inputs
• Reduced manual annotation time significantly
• Ensured consistency and reproducibility across datasets

Use Case & Impact

• Enables rapid generation of training data for computer vision models
• Eliminates dependency on real-world annotated datasets
• Bridges 3D content creation workflows with AI training pipelines

‍