The Convergence Problem

The Game Ends, The Work Begins

The credits rolled on Alan Wake 2. For a moment, I sat in that holy silence after a story ends.

Then reality returned. Homework due tomorrow.

Stories have a way of getting under your skin. They reframe every equation as a metaphor, turn every problem into a reminder of what you’re trying not to think about.

I opened the PDF. Partial differential equations stared back at me, incomprehensible as grief.

∂u/∂t = α∇²u

The heat equation. How warmth dissipates through space and time. How things fade away.

Like her.

My computer vision code waited—teaching machines to detect edges, to find boundaries between light and dark, between what’s there and what’s not.

Between who I was with you and who I am without you.

Alan Wake wrote his way through darkness to return to Alice. I was just trying to survive until tomorrow, one equation at a time.

In the darkness, something whispered: “It’s not a lake. It’s an ocean.”

The homework would get done. The missing wouldn’t.

That was my story now.

[END]

🚗 Vehicle Submersion Detection System

水积识别和车辆淹没部位判别系统

A comprehensive system for detecting vehicles in flood images and analyzing their submersion levels using custom-trained YOLO models.

---

📋 Overview

This system uses custom-trained YOLO models to:

• Detect vehicle parts (body, window, wheel) in flood images
• Segment water areas using semantic segmentation
• Calculate submersion levels by analyzing part-water overlap
• Generate statistics and color-coded visualizations

Built for academic research on flood disaster assessment and vehicle damage analysis.

---

✨ Features

🎯 Part-Based Vehicle Detection

• Detects vehicle components separately: body, window, wheel
• Custom YOLO model trained on 100 epochs (exp23)
• Confidence threshold: 10% (adjustable 1-100%)

🌊 Water Segmentation

• Semantic segmentation for water area identification
• Custom YOLO segmentation model
• 60% opacity bright blue overlay for clear visualization

📊 Submersion Analysis

• Fully Submerged: ≥70% of body/window underwater
• Partially Submerged: Wheels underwater or 20-70% body coverage
• Not Submerged: <20% water coverage
• Per-part submersion ratio calculation

🎨 Interactive Web Interface

• Built with Streamlit for easy interaction
• Multi-model selection (custom + official YOLO models)
• Real-time processing with progress indicators
• Three view modes: Detection Only, Segmentation Only, Combined
• Split-screen comparison (original vs processed)

📥 Export & Download

• Download all results as ZIP archive
• Individual downloads per image (detection, segmentation, combined, stats JSON)
• Aggregate statistics across multiple images
• CSV export support

---

🚀 Quick Start

Prerequisites

# Conda environment with Python 3.8+
conda activate yolov11

# Required packages
pip install -r requirements.txt

Run the Application

# Method 1: Direct command
streamlit run app.py

# Method 2: Batch file (Windows)
run.bat

The web interface will open at http://localhost:8501

---

📁 Project Structure

Project3/
├── README.md                   # This file
├── GUIDE.md                    # Quick start guide
├── run.bat                     # Application launcher
├── requirements.txt            # Python dependencies
│
├── Core Application
│   ├── app.py                  # Streamlit web interface
│   ├── detection.py            # Vehicle part detection module
│   ├── segmentation.py         # Water segmentation module
│   ├── utils.py                # Analysis and visualization
│   └── config.py               # Configuration and model paths
│
├── models/                     # Trained YOLO models
│   ├── vehical/
│   │   ├── best_exp23.pt      # Best vehicle detection (12.37% max conf)
│   │   ├── last_exp23.pt      # Last epoch checkpoint
│   │   ├── best.pt            # Improved model (0.44% max conf)
│   │   └── last.pt
│   └── flood/
│       ├── best.pt            # Best water segmentation
│       └── last.pt            # Last epoch checkpoint
│
├── demo_images/                # Sample flood images
├── results/                    # Processing outputs
├── uploads/                    # User uploaded images
│
├── README/                     # Documentation
│   ├── README.md              # Complete documentation
│   ├── requirement.md         # Project requirements
│   ├── project_details.md     # Implementation details
│   ├── architecture.md        # System architecture
│   ├── CUSTOM_MODELS.md       # Custom model info
│   ├── CUSTOM_MODEL_CLASSES.md # Class definitions
│   └── SETUP_COMPLETE.md      # Setup guide
│
└── test/                       # Test scripts (19 files)
    ├── test_flood_now.bat/.py
    ├── test_exp23.bat/.py
    ├── test_custom.bat/.py
    └── ...

---

🎯 Usage Guide

1. Launch Application

streamlit run app.py

2. Configure Models (Sidebar)

• Detection Model: Select vehicle detection model
  • Recommended: 🎯 Vehicle Custom (Exp23 Best)
• Segmentation Model: Select water segmentation model
  • Recommended: 🌊 Flood Custom (Best)
• Confidence Threshold: Adjust sensitivity (default: 0.10)

3. Process Images (Tab 1)

1. Click "Browse files" to upload images
2. Select one or multiple images (JPG, PNG, BMP)
3. Click "🚀 Process Images"
4. View results with radio buttons:
   • Detection Only: Vehicle bounding boxes only
   • Segmentation Only: Water segmentation only
   • Combined: Both detection and segmentation

4. View Statistics (Tab 2)

• Overall statistics across all images
• Per-image breakdown table
• Submersion level distribution chart

5. Download Results (Tab 3)

• Download all results as ZIP
• Or download individual results per image

---

🧪 Testing

Test scripts are located in the test/ directory:

# Test flood segmentation
cd test
test_flood_now.bat

# Test exp23 vehicle detection
test_exp23.bat

# Test all images in demo_images/
test_all.bat

# Quick single-image test
quick_test.bat

---

🔧 Configuration

Edit config.py to customize:

# Model paths
DETECTION_MODELS = {
    "🎯 Vehicle Custom (Exp23 Best)": "models/vehical/best_exp23.pt",
    # Add more models...
}

SEGMENTATION_MODELS = {
    "🌊 Flood Custom (Best)": "models/flood/best.pt",
    # Add more models...
}

# Custom class definitions
CUSTOM_VEHICLE_PARTS = {0: 'body', 1: 'window', 2: 'wheel'}
CUSTOM_WATER_CLASSES = {0: 'water'}

---

📊 Technical Details

Models

Vehicle Detection Model (Exp23)

• Architecture: YOLOv8/v11 object detection
• Classes: body (0), window (1), wheel (2)
• Training: 100 epochs on project1 dataset
• Performance: 12.37% max confidence on test set
• Input: Ground-level vehicle images

Water Segmentation Model

• Architecture: YOLOv8/v11 semantic segmentation
• Classes: water (0)
• Training: Custom flood dataset
• Output: Binary mask for water regions

Submersion Analysis Algorithm

# Whole vehicle submersion
if submersion_ratio >= 0.7:
    level = "fully_submerged"
elif submersion_ratio >= 0.2:
    level = "partially_submerged"
else:
    level = "not_submerged"

# Part-based submersion (custom model)
if part == 'wheel' and submersion_ratio >= 0.5:
    level = "partially_submerged"
elif part in ['body', 'window'] and submersion_ratio >= 0.7:
    level = "fully_submerged"

Visualization

• Red: Fully Submerged (≥70%)
• Orange: Partially Submerged (20-70%)
• Green: Not Submerged (<20%)
• Blue: Water Area (60% opacity overlay)

---

⚠️ Important Notes

Model Limitations

The custom vehicle detection model was trained on ground-level clear vehicle photos, NOT aerial flood images. For optimal results:

✅ Best: Use images from training dataset

• Location: 机器学习课程实践\project1\dataset2\images\test\

⚠️ Limited: New aerial flood images (cc*.jpg)

• Detection may be poor due to angle/lighting differences
• Consider retraining with flood-specific images

🔄 Alternative: Use official YOLO models (YOLOv11n) for general vehicles

Confidence Threshold

• Default: 0.10 (10%)
• Range: 0.01 to 1.0
• Lower = more detections (more false positives)
• Higher = fewer detections (higher precision)

---

🐛 Troubleshooting

No Detections?

1. Lower confidence threshold (try 0.05)
2. Use images from training dataset
3. Check console output for model loading confirmation

Water Not Visible?

1. Water overlay is 60% opacity with bright blue
2. Check console for segmentation results
3. Verify water exists in image

Results Disappear When Switching Views?

• Fixed: Results now persist across view switches
• Ensure you're in "Upload & Process" tab

---

📚 Documentation

Detailed documentation available in README/:

• README.md: Complete system documentation
• GUIDE.md: Quick start guide
• requirement.md: Original requirements
• architecture.md: Technical architecture
• CUSTOM_MODELS.md: Model information
• project_details.md: Implementation notes

---

🛠️ Technology Stack

• Frontend: Streamlit 1.28+
• Backend: Python 3.8+
• Deep Learning: PyTorch 2.0+ with CUDA
• Computer Vision: OpenCV 4.8+
• Object Detection: Ultralytics YOLO (v8 & v11)
• Data Processing: NumPy, Pandas
• Image Handling: PIL/Pillow

---

📦 Dependencies

streamlit>=1.28.0
ultralytics>=8.0.0
opencv-python>=4.8.0
torch>=2.0.0
torchvision>=0.15.0
numpy>=1.24.0
pandas>=2.0.0
Pillow>=10.0.0

Install all:

pip install -r requirements.txt

---

🎓 Academic Context

Project: Task 4 - Vehicle Submersion Detection System Course: Machine Learning Practice (机器学习课程实践) Institution: [Your Institution] Due Date: October 28, 2025 Environment: Conda (yolov11), Python 3.x

---

📝 License

This project is for academic use. Please refer to your institution's policies for code sharing and reuse.

---

🤝 Contributing

This is an academic project. For questions or issues:

1. Check README/ documentation
2. Review test/ scripts for examples
3. Consult GUIDE.md for troubleshooting

---

📧 Contact

For questions related to this project, please contact the development team.

---

🙏 Acknowledgments

• Ultralytics YOLO: Object detection framework
• Streamlit: Web interface framework
• OpenCV: Image processing library
• PyTorch: Deep learning platform

---

📊 Project Statistics

• Lines of Code: ~2,000+
• Python Files: 5 core modules
• Test Scripts: 19 test files
• Documentation: 7 markdown files
• Models: 6 trained models (2 best + 4 checkpoints)

---

⭐ Star this repository if you find it helpful!

🚀 Get started: streamlit run app.py "# -yolo-promax"