GroundingDINO

GroundingDINO

GroundingDINO by IDEA Research is an open-vocabulary object detection model that fuses the DINO detector with BERT-based text grounding. You describe what you want to find in plain English and the model localises it, making it ideal for zero-shot detection pipelines and as the vision backbone for agents.

Key Features

• Open-set detection: detect any category described in natural language
• State-of-the-art zero-shot COCO performance (49.8 AP) without COCO training
• Seamless integration with SAM for open-vocabulary instance segmentation (Grounded-SAM)
• Feature-enhanced transformer with cross-attention between text and image features
• Pretrained weights for both Swin-T and Swin-B backbones

Quick Start

pip install groundingdino-py

from groundingdino.util.inference import load_model, load_image, predict

model = load_model("groundingdino_swint_ogc.py", "groundingdino_swint_ogc.pth")
image_source, image = load_image("image.jpg")
boxes, logits, phrases = predict(
    model=model, image=image,
    caption="a cat . a dog . a person",
    box_threshold=0.35, text_threshold=0.25,
)
print(phrases)

npx ai-supply add groundingdino-open-vocabulary-detection

Curated mirror of the open-source GroundingDINO (Apache-2.0). Get it from the source.