metadata
license: mit
library_name: ultralytics
pipeline_tag: object-detection
tags:
- building-damage-detection
- disaster-response
- computer-vision
- tornado
TornadoNet: Real-Time Building Damage Detection with Ordinal Supervision
TornadoNet is a comprehensive benchmark for automated street-level building damage assessment following tornado events. It evaluates modern real-time object detection architectures, comparing CNN-based YOLO models with transformer-based RT-DETR models. The project introduces ordinal-aware supervision strategies to improve multi-level damage classification (DS0-DS4) under realistic post-disaster conditions.
- Paper: TornadoNet: Real-Time Building Damage Detection with Ordinal Supervision
- GitHub Repository: crumeike/TornadoNet
- Dataset: tornadonet-datasets
Key Features
- Standardized Benchmark: Evaluates YOLOv8, YOLO11, and RT-DETR using 3,333 high-resolution images and 8,890 building instances from the 2021 Midwest tornado outbreak.
- Ordinal Supervision: Introduces soft ordinal targets and distance penalties, improving damage severity estimation by 4.8% mAP for RT-DETR models.
- Real-Time Performance: Optimized for high throughput, with variants reaching up to 276 FPS on NVIDIA A100 GPUs.
- Five-Level Classification: Based on the IN-CORE damage states (DS0 to DS4).
Sample Usage
This model can be used with the ultralytics library.
Installation
pip install ultralytics huggingface_hub
Inference
from ultralytics import YOLO
from huggingface_hub import hf_hub_download
# Download a specific model checkpoint (e.g., YOLO11-x baseline)
model_path = hf_hub_download(
repo_id="crumeike/tornadonet-checkpoints",
filename="tornadonet-yolo11-x-baseline/best.pt"
)
# Load model
model = YOLO(model_path)
# Run inference
results = model.predict(
source="path/to/image.jpg",
imgsz=896,
conf=0.25
)
# Visualize results
results[0].show()
Damage Classification Levels
The model predicts damage states according to the following framework:
| Class | Label | Description |
|---|---|---|
| DS0 | Undamaged | No visible damage |
| DS1 | Slight | Minor roof/window damage |
| DS2 | Moderate | Significant roof damage |
| DS3 | Extensive | Major structural damage |
| DS4 | Complete | Total collapse |
Citation
@article{umeike2026tornadonet,
title={TornadoNet: Real-Time Building Damage Detection with Ordinal Supervision},
author={Umeike, Robinson and Pham, Cuong and Hausen, Ryan and Dao, Thang and Crawford, Shane and Brown-Giammanco, Tanya and Lemson, Gerard and van de Lindt, John and Johnston, Blythe and Mitschang, Arik and Do, Trung},
journal={arXiv preprint arXiv:2603.11557},
year={2026}
}