HDINO

This is the official implementation of HDINO.

Please feel free to reach out if you have any questions or suggestions.

Updates

• [2026-03-13]: Repository initialized and model weights uploaded.

Architecture

Zero-shot Results on COCO 2017 val

Model	Backbone	Training Data	Zero-Shot 2017val
DyHead-T	Swin-T	O365	43.6
GLIP-T (B)	Swin-T	O365	44.9
GLIP-L	Swin-L	FourODs, GoldG, Cap24M	49.8
Grounding-DINO-T	Swin-T	O365, GoldG, Cap4M	48.4
Grounding-DINO-L	Swin-L	O365, OpenImages, GoldG	52.5
T-Rex2-T	Swin-T	O365, GoldG, OpenImages, Bamboo, CC3M, LAION	46.4
YOLO-World-S	YOLOv8-S	O365, GoldG	37.6
YOLO-World-M	YOLOv8-M	O365, GoldG	42.8
YOLO-World-L	YOLOv8-L	O365, GoldG	44.4
YOLO-World-L	YOLOv8-L	O365, GoldG, CC3M	45.1
HDINO-T (ours)	Swin-T	O365, OpenImages	49.2
HDINO-L (ours)	Swin-L	O365, OpenImages	51.7

1. Create a virtual environment

conda create -n hdino python=3.10 -y
conda activate hdino

2. Install PyTorch and CUDA

The following command installs the specific versions used in our development environment (PyTorch 2.1.0 + CUDA 12.1):

torch           2.1.0+cu121
torchvision     0.16.0+cu121

3. Install custom operators

Navigate to the models/HDINO/ops directory and install the operators:

cd models/HDINO/ops
pip install -e .

🎨 Demo

You can run our interactive demo locally to experience HDINO:

python gradio_demo.py

📜 Acknowledgement

We express our sincere gratitude to the authors for their contributions to the community:

• DINO
• Grounding DINO
• Open-GroundingDino