Skip to main content

Grounding DINO Tiny

Grounding DINO Tiny is a smaller, cheaper open-vocabulary detector. You give it an image and a text prompt like “signature”, “invoice total”, or “person wearing helmet”, and it returns bounding boxes.

Learning goals​

  • Understand when “Tiny” is good enough (speed/cost vs accuracy)
  • Build a reliable post-processing pipeline (thresholds + NMS)
  • Serve detections behind a simple API

Why “open-vocabulary” matters​

Instead of training a custom detector for every label, you can:

  • start with a text prompt
  • iterate quickly on real data
  • add a finetuned model later if needed

Output format (what you should standardize)​

For production pipelines, standardize on one JSON format:

json
{
  "prompt": "signature",
  "boxes": [
    {"x1": 120, "y1": 340, "x2": 410, "y2": 515, "score": 0.62}
  ]
}

Post-processing checklist​

  • Confidence threshold (e.g., 0.25–0.5)
  • NMS (remove overlapping boxes)
  • Clamp boxes to image boundaries
  • Convert to a consistent coordinate system (pixel vs normalized)
Rule of thumb

If you’ll crop regions (like signatures), err on the side of slightly larger boxes to avoid cutting off content.

Packaging as an API​

A simple API contract for a detection service:

  • POST /detect with image + prompt
  • returns boxes + scores
  • optional POST /crop returns cropped PNGs

This is exactly what Lab 5 (Signature Detection) builds toward.

Where this fits​

Start with Grounding DINO for fundamentals, then use Tiny when latency/cost matters.