Aerial detection of HVAC systems (Case Study for Grande Data Services)

Contains:

• Assessment of Grande Data labeler performance
• Training of a custom object detection model using GroundingDino as a base
  • Text prompt tuning based on multi-run zero-shot performance and automatic LLM (ChatGPT) prompt reselection
  • Object detection reranking based on supervised learning from the zero-shot last hidden layer

Results:

• Dramatically improved object detection performance of rooftop HVAC systems based training from only 120 labeled images