corzent/yolov12x-dino3-watermark-detection

🔍 YOLOv12x-DINOv3 Watermark Detection Model

<div align="center">

   

State-of-the-art watermark detection powered by YOLOv12 + DINOv3 Vision Transformers

</div>

📊 Model Performance

🏗️ Architecture

• —Base Model: YOLOv12x (Extra-Large variant)
• —Enhancement: DINOv3 ViT-B/16 backbone integration
• —Configuration: Dual P0/P3 feature enhancement
• —Input Size: 1024×1024

Key Features

• —🧠 DINOv3 Vision Transformer integration at P4 level (40×40×512)
• —🔄 Dual-scale feature fusion for improved small/medium object detection
• —⚡ Optimized for watermark detection with high precision
• —🛡️ Production-ready with comprehensive error handling

🚀 Quick Start

Installation

pip install ultralytics

Inference

from ultralytics import YOLO

# Load model from Hugging Face
model = YOLO('hf://corzent/yolov12x-dino3-watermark-detection')

# Or load locally
model = YOLO('best.pt')

# Run inference
results = model('image.jpg')

# Process results
for result in results:
    boxes = result.boxes
    for box in boxes:
        print(f"Watermark detected: confidence {box.conf[0]:.2f}")

Batch Processing

# Process multiple images
results = model(['image1.jpg', 'image2.jpg', 'image3.jpg'])

# Process video
results = model('video.mp4', stream=True)

📋 Training Details

📈 Training Progress

The model was trained for 100 epochs with consistent improvement:

• —Epoch 1: mAP50 = 0.3%
• —Epoch 25: mAP50 = 53.7%
• —Epoch 50: mAP50 = 77.2%
• —Epoch 75: mAP50 = 79.6%
• —Epoch 100: mAP50 = 84.1%

🔧 Model Configuration

The model uses the yolov12x-dino3-vitb16-dual.yaml configuration:

# YOLOv12x with DINOv3 ViT-B/16 Dual Enhancement
backbone:
  - YOLOv12x backbone layers
  - DINOv3 ViT-B/16 integration at P4

head:
  - Multi-scale detection head
  - Dual P0/P3 feature enhancement

📁 Files Included

• —best.pt - Best model checkpoint (highest mAP)
• —last.pt - Final epoch checkpoint
• —args.yaml - Training configuration
• —results.csv - Training metrics log
• —*.png - Training curves and confusion matrices

⚠️ Limitations

• —Optimized specifically for watermark detection
• —Requires GPU with 8GB+ VRAM for inference at 1024×1024
• —Best performance on images similar to training distribution

📄 License

This model is released under the AGPL-3.0 License.

🙏 Acknowledgments

• —**Sompote/DINOV3-YOLOV12** - Original YOLOv12 + DINOv3 integration
• —Ultralytics - YOLO framework
• —Meta AI - DINOv2/v3 Vision Transformers
• —PyTorch - Deep learning framework

📞 Citation

If you use this model, please cite the original YOLOv12-DINOv3 repository:

@software{sompote_yolov12_dinov3_2024,
  title={YOLOv12 + DINOv3 Vision Transformers Integration},
  author={Sompote},
  year={2024},
  url={https://github.com/Sompote/DINOV3-YOLOV12}
}