P-001 · 2025 · Hackathon
Broken-tile detection with YOLOv8
A 24-hour ChefTreff AI Hackathon prototype: a YOLOv8 model detects broken roof tiles in photos, served through a FastAPI upload endpoint and a web interface.
From site photo to prediction in 24 hours
- PreparationImage collection and preprocessing notebooks
- AnnotationCVAT labels exported in YOLO format
- Detection modelUltralytics YOLOv8 trained on broken tiles
- ApplicationFastAPI endpoint and HTML client in Docker
A user uploads a photo through the web client
The challenge
The project explored how images taken on a construction site could help identify damaged roof tiles. The task was to detect visible defects in an image and make the prediction accessible through a small application, within the 24-hour hackathon.
Data and model
The workflow includes image collection and preprocessing notebooks, annotation in CVAT and labels exported in YOLO format. An Ultralytics YOLOv8 model was trained to detect broken tiles. The repository includes the trained checkpoint and prediction outputs.
From model to application
A FastAPI endpoint accepts an uploaded image, converts it to RGB and runs the detector. The implemented response reports whether a broken tile was detected and how many detections were returned. A lightweight HTML client provides the upload interface, while Docker packages the application and its dependencies.
What the prototype demonstrates
The project connects data preparation, object detection and an application interface in one workflow. It is a hackathon prototype: the presence of a trained model and demo does not establish readiness for construction-site use. Broader validation across lighting conditions, viewpoints and types of damage would be needed.