DAM

A powerful object detection system built on YOLO (You Only Look Once) architecture, specifically fine-tuned for the LVIS (Large Vocabulary Instance Segmentation) dataset. This project provides both training and inference capabilities, along with a user-friendly Streamlit web interface for real-time object detection. The system is specifically designed to handle the challenges of large-scale object detection with many classes and imbalanced datasets.

About DAM-1

DAM-1 (Detect Anything Model v1) is a breakthrough in large-scale object detection, capable of identifying objects across 1200 different classes. What makes this achievement even more remarkable is that it was entirely trained by the Hexo MLE agent - an autonomous AI system that successfully tackled one of the most challenging tasks in computer vision: training a large-scale object detection model to convergence.

To put this in perspective, Mediapipe - one of the most commonly used object detection models - can only predict 80 categories of objects. DAM-1 represents a 15x increase in detection capabilities, demonstrating the power of autonomous AI systems in solving complex machine learning challenges.

Learn more about the Hexo Agent and its capabilities at hexo.ai

Training Challenges

Training large-scale object detection models presents several significant challenges:

• Class Imbalance: The LVIS dataset exhibits severe class imbalance, with some classes having thousands of instances while others have only a few. This imbalance makes it difficult for the model to learn rare classes effectively.

• Convergence Issues: With 1203 different classes, achieving convergence is particularly challenging. The model needs to learn discriminative features for each class while maintaining good performance across all classes.

• Computational Complexity: Training on such a large number of classes requires significant computational resources and careful optimization of the training process.

• Memory Constraints: Handling a large number of classes and instances requires efficient memory management and optimization techniques.

Our implementation addresses these challenges through:

• Advanced sampling strategies to handle class imbalance
• Careful learning rate scheduling
• Gradient accumulation for effective batch processing
• Regularization techniques to prevent overfitting
• Progressive training approaches

Features

• 🎯 Large-Scale Detection: YOLO-11n model fine-tuned for LVIS dataset with 1200+ object classes - 15x more categories than Mediapipe
• 🔍 Comprehensive Coverage: Detect virtually any object in your images with high accuracy
• 📊 Interactive web interface using Streamlit
• 🔄 Support for single image and batch processing
• 📈 Comprehensive evaluation metrics
• 🖼️ Real-time object detection with confidence scores
• 📝 Detailed detection results with bounding boxes
• 🚀 Autonomous Training: Successfully trained by Hexo MLE agent, demonstrating advanced AI capabilities

Installation

1. Clone the repository:

git clone https://github.com/hexo-ai/dam.git
cd dam

2. Install dependencies:

pip install -r requirements.txt

Usage

Web Interface

Launch the Streamlit web interface:

streamlit run streamlit_app.py

The web interface provides three main features:

1. Single Image Prediction
2. Batch Prediction
3. Test Set Prediction

Training

To train the model on your dataset:

python train.py

Evaluation

To evaluate the model's performance:

python evaluate.py

Project Structure

dam/
├── streamlit_app.py    # Web interface for object detection
├── train.py           # Training script
├── evaluate.py        # Evaluation script
├── requirements.txt   # Project dependencies
└── documentations/    # Additional documentation

Dependencies

Key dependencies include:

• PyTorch
• Ultralytics YOLO
• Streamlit
• OpenCV
• NumPy
• Pandas

For a complete list of dependencies, see requirements.txt.

Model Details

The project uses YOLO-11n, a state-of-the-art object detection model, fine-tuned on the LVIS dataset. The model can detect objects from 1203 different classes.

Performance Comparison

The assets folder contains visual comparisons between DAM and Mediapipe detection results, along with validation metrics:

Validation Performance

Detection Comparisons

Here are side-by-side comparisons of object detection results between DAM and Mediapipe, demonstrating the superior performance of our model in various scenarios:

These comparisons highlight DAM's improved accuracy and robustness in object detection tasks compared to Mediapipe's baseline implementation.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Acknowledgments

• Ultralytics for the YOLO implementation
• LVIS dataset team