Yolo-DinoV2

Yolo-DinoV2 integrates Meta's DINOv2, a self-supervised Vision Transformer model, as a frozen feature extractor backbone into Ultralytics' YOLO object detection framework. This combination aims to enhance few-shot object detection capabilities, allowing for effective training with minimal data.

Features

• ⭐ DINOv2 Backbone: Utilizes DINOv2's robust feature extraction to improve object detection performance. All DinoV2 backbone sizes supported (uses registers)

• ⭐ Seamless Integration: Maintains compatibility with Ultralytics' YOLO, enabling standard training and inference workflows.

• ⭐ Few-Shot Learning: Designed to generalize well with relatively small datasets, facilitating quick training on custom data.

Installation

1. Clone the Repository:

   git clone https://github.com/itsprakhar/Yolo-DinoV2.git
   cd Yolo-DinoV2

2. Install Dependencies:

   Ensure you have Python installed, then run:

   pip install -r requirements.txt

Usage

1. ⚙️ Model Initialization:

   Use the configuration files in the yolo_dinov2_configs directory to initialize the YOLO model with the DINOv2 backbone. Modify the nc parameter in the YAML file to match the number of classes in your dataset.

2. 🏋️ Training:

   Train the model as you would with any YOLO model from Ultralytics. Refer to Ultralytics' training documentation for detailed instructions.

3. 🏃Inference:

   Same you would with any YOLO model from Ultralytics. Refer to Ultralytics' Inference documentation for detailed instructions.

• 🥹 Pretrained Weights: Pretrained weights are not available. However, the model can be trained quickly on custom data with relatively small datasets 😊, thanks to its strong generalization capabilities.

Contributing

Contributions are welcome! Feel free to open issues or submit pull requests to enhance the functionality of this project.

Acknowledgements

• DINOv2 🦕: Developed by Meta AI Research, DINOv2 is a self-supervised Vision Transformer model that produces high-performance visual features applicable across various computer vision tasks.

• Ultralytics YOLO 🚀: A state-of-the-art real-time object detection model known for its speed and accuracy.

---

If you find this project useful, please consider giving it a ⭐ on GitHub to motivate further development.

Enjoy using Yolo-DinoV2 for your object detection tasks! 🤩