HOMEY

🏠 HOMEY: Heuristic Object Masking with Enhanced YOLO for Property Insurance Risk Detection

Teerapong Panboonyuen
MARSAIL (Motor AI Recognition Solution Artificial Intelligence Laboratory)
Preprint on arXiv (March 2026) 📄
arXiv: 2603.18502

HOMEY Architecture

🚀 Overview

HOMEY is a novel framework for AI-driven property risk detection, combining YOLO object detection with domain-specific heuristic masking and risk-aware loss calibration.

The model targets 17 classes of property risk indicators, including:

Structural damage (cracked foundations, roof issues)
Maintenance neglect (dead yards, overgrown bushes)
Liability hazards (falling gutters, hazard signs, garbage)

Property imagery often contains cluttered scenes with subtle risk cues. HOMEY enhances detection by amplifying weak signals through heuristic object masking while maintaining fast, reliable inference, suitable for real-world insurance workflows.

✨ Key Innovations

🔹 Heuristic Object Masking

Enhances weak visual signals in complex, cluttered environments
Focuses YOLO on regions most indicative of risk

🔹 Risk-Aware Loss Calibration

Balances class skew
Weights severe hazards higher than minor issues
Improves model reliability for insurance-critical tasks

🔹 Integrated YOLO Enhancement

Maintains real-time inference speeds
Compatible with YOLOv8 backbones
Easily extendable to new property risk classes

📊 Qualitative Results — Risk Highlighting

Qualitative Detection Results

Models without heuristic masking often miss subtle hazards or over-focus on background clutter. HOMEY shifts attention to high-risk property regions, improving interpretability for underwriting teams.

📈 Detection Performance

Performance Metrics

On real-world property datasets:

✅ Higher mAP (mean Average Precision) across 17 classes ✅ Reduced false negatives for critical hazards ✅ Fast inference (~YOLO-level speed) ✅ Stable performance under varied property conditions

📖 Official Publication

🏆 arXiv Preprint (March 2026)

🔗 arXiv Link: https://arxiv.org/abs/2603.18502

📌 DOI: 10.48550/arXiv.2603.18502

🧠 Why HOMEY Matters

Property risk detection for insurance requires models that are:

✔ Robust to cluttered scenes and variable image quality
✔ Sensitive to both subtle and severe hazards
✔ Interpretable for underwriting decisions
✔ Fast enough for real-time inspection workflows

HOMEY achieves this by enhancing signal extraction rather than just increasing network depth or dataset size.

🚀 Training HOMEY

We provide a reproducible PyTorch pipeline in the src/ directory. HOMEY supports:

🧠 YOLOv8 backbones
🏠 Multi-class property risk training (17 risk classes)
🔍 Heuristic object masking integration
📈 Risk-aware loss weighting
💻 GPU-accelerated training

📂 Project Structure

HOMEY/
│
├── src/
│   ├── train_homey.py
│   ├── models/
│   ├── datasets/
│   ├── losses/
│   └── utils/
│
├── data/
│   ├── train/
│   ├── val/
│   └── test/
│
└── outputs/

Each class folder in train/ or val/ should contain images labeled by risk type:

train/
    cracked_foundation/
    overgrown_bushes/
    falling_gutters/
    ...

⚙️ Installation

git clone https://github.com/kaopanboonyuen/HOMEY.git
cd HOMEY

conda create -n homey python=3.10
conda activate homey

pip install -r requirements.txt

▶️ Training Example

python src/train_homey.py \
    --train_dir data/train \
    --val_dir data/val \
    --backbone yolov8n \
    --epochs 100 \
    --batch_size 16 \
    --lr 3e-4

💾 Outputs

outputs/
    best_model.pth
    training_logs/
    detection_results/

Automatic tracking of:

✅ Validation mAP per class ✅ Risk-weighted loss curves ✅ Checkpoint saving with GPU support

🧠 Why This Training Matters

HOMEY training directly targets real-world insurance risk detection:

Suppresses false alarms in cluttered backgrounds
Enhances detection of subtle hazard cues
Balances minor vs. severe risks with custom loss
Produces interpretable detection masks for human review

The model sees property risk — not irrelevant background clutter.

🙏 Acknowledgement

🎓 Talent Scholarship for Exceptional Ability 🏫 College of Computing, Khon Kaen University

📚 BibTeX Citation

@article{panboonyuen2026homey,
  author    = {Teerapong Panboonyuen},
  title     = {HOMEY: Heuristic Object Masking with Enhanced YOLO for Property Insurance Risk Detection},
  journal   = {arXiv preprint arXiv:2603.18502},
  year      = {2026},
  doi       = {10.48550/arXiv.2603.18502}
}

⭐ If you find this project useful, please consider starring the repository!

This site is open source. Improve this page.