🔍 A Submission for the Duality AI - Space Station Hackathon

EVA Guardian is an AI-powered safety monitoring system designed to enhance operational safety and situational awareness aboard space stations.

At its core is a YOLOv8 object detection model, trained on a fully synthetic dataset from Duality AI’s Falcon platform. The model detects three mission-critical assets:

• 🔥 Fire Extinguishers
• 🧰 Toolboxes
• 🧪 Oxygen Tanks

This end-to-end solution includes a training pipeline, feedback loop, and a Streamlit dashboard for live image/webcam detection and safety reports.

📈 Our final model achieved an impressive mAP@0.5 of 0.936, proving the power of synthetic data for real-world applications.

• Model: YOLOv8 (ultralytics)
• Frameworks: PyTorch, OpenCV, NumPy, PIL
• Frontend: Streamlit, Plotly
• Environment: Python 3.10 (Conda)
• Data Platform: Duality AI Falcon (Synthetic)

• ✅ High-Performance Detection: YOLOv8 trained for 50 epochs with augmentations (mosaic, mixup, etc.). Final metrics show strong generalization to unseen data.

• 🖥️ Interactive Dashboard (Streamlit): Upload images or use your webcam to run real-time object detection. Automatically loads latest trained model.

• 📋 Actionable Safety Report: Color-coded alerts (CRITICAL, WARNING, OK) based on object detection — tailored for space safety applications.

• 🔄 Human-in-the-Loop Feedback: Users can draw boxes over missed objects. A feedback processor script converts annotations into retrainable data.

• 📊 Training Metrics Viewer: Expandable section to view confusion matrix, loss curves, and precision-recall charts — all generated from training.

Trained for 50 epochs on synthetic data with custom YOLO parameters.

Prerequisites:

• Anaconda / Miniconda
• Python 3.10
• NVIDIA GPU with CUDA (for training)

# Clone the repo
git clone https://github.com/your-username/eva-guardian.git
cd eva-guardian

# Create and activate the environment
conda create -n duality_ai python=3.10
conda activate duality_ai

# Install dependencies
pip install -r requirements.txt

streamlit run app.py

The app auto-loads your latest YOLO model from the runs/ directory. Supports image upload, webcam, and drawing feedback boxes.

If you’ve collected new feedback using the app and processed it using process_feedback.py, you can re-train the model:

python train.py --resume

You can also use the improved train_resumable.py to automatically save progress, resume training, and avoid data loss.

• ✅ Automated Feedback Loop: Auto-retraining using a monitoring service for feedback.json.

• ✅ Precise JS-Based Annotation: Integrate frontend JS tools for exact coordinate capture.

• ✅ Advanced Synthetic Generation: Use domain randomization via Duality Falcon to improve sim-to-real generalization.

.
├── data/                  # Dataset (excluded from GitHub)
├── new_user_images/       # Processed feedback images
├── runs/                  # YOLO training outputs (excluded from GitHub)
├── app.py                 # Streamlit dashboard
├── train.py               # Main training script
├── train_resumable.py     # Auto-resumable training version 
├── predict.py             # Inference script
├── process_feedback.py    # Converts user boxes to YOLO format
├── yolo_params.yaml       # YOLOv8 config
├── requirements.txt       # Dependencies
├── Final_Report.pdf       # Submission report
└── README.md              # This file