beaunix/smart-crack-vision

SmartCrackLens — Crack Detection & Severity Classification

This repository hosts the vision model that powers SmartCrackLens, an agent for crack detection and severity classification on structural surfaces (roads, walls, infrastructure).

Built with Ultralytics YOLOv8. Full credit to Ultralytics for the open YOLO architecture and training tooling.

Model

• Architecture: YOLOv8-nano (instance segmentation)
• Task: Single-class (crack) instance segmentation
• Formats provided:
  • smartcracklens-seg.pt — Ultralytics checkpoint. Use for fine-tuning or retraining.
  • smartcracklens-seg.onnx — ONNX export. Use for inference in any ONNX Runtime environment.

Training data

Trained on the Crack Segmentation Dataset, officially distributed by Ultralytics as part of their datasets collection (4,029 annotated road/wall crack images, single crack class). Originally sourced from Roboflow Universe and marked Public Domain by Ultralytics.

Citation (as given by Ultralytics): ```bibtex @misc{ crack-bphdr_dataset, title = { crack Dataset }, type = { Open Source Dataset }, author = { University }, url = { https://universe.roboflow.com/university-bswxt/crack-bphdr }, year = { 2022 }, month = { dec }, note = { visited on 2024-01-23 }, } ```

Usage

ONNX Runtime (inference only)

```python import onnxruntime as ort

session = ort.InferenceSession("smartcracklens-seg.onnx") outputs = session.run(None, {"images": input_tensor}) ```

Ultralytics (.pt — inference or further training)

```python from ultralytics import YOLO

model = YOLO("smartcracklens-seg.pt") results = model.predict("wall_photo.jpg") ```

License

Released under AGPL-3.0, in line with the Ultralytics YOLO license this model was trained with (no Enterprise license was purchased). The training data itself is Public Domain.

Related

• Project repo: https://github.com/BeauBryanDev/SmartCrackLens