Istanbul Sewer Defect Detection Model (ONNX)
This repository contains the trained weights of our deep learning models for automated sewer defect detection.
The models were trained on the Istanbul Water and Sewerage Administration (İSKİ) Sewer Defect Dataset, developed in collaboration with the İSKİ.
Due to data privacy restrictions, the raw images are not publicly available, but the models and metadata are released for research purposes.
However, the permission process for data release has been initiated, and if approved by the General Directorate, the dataset will be published on the Hugging Face platform for research use.
📄
Data Access Policy:
A
Data Request Form has been prepared for researchers.
Upon completing this form and submitting an official request via
İSKİ’s institutional KEP (Registered Electronic Mail) address
📨
iskigenelmudurlugu@hs01.kep.tr
The
raw images used in this thesis may be shared for
non-commercial academic research purposes,
subject to ISKI’s review and approval in compliance with data protection regulations.
License
These models are released under the Fair Non-Commercial Research License.
They can be freely used for academic and research purposes only.
Commercial use is not permitted.
🧩 Dataset Description
The Istanbul Sewer Defect Dataset (ISWDS) is a novel, expert-curated benchmark designed to advance automated sewer defect detection.
The imagery was collected between 2021–2024 across all 39 districts of Istanbul, using robotic CCTV systems under diverse operational conditions (lighting, flow, turbidity, and camera types).
All videos were recorded in high-quality and frames were extracted and manually filtered to remove blur or poor visibility.
This dataset is based on the EN 13508-2:2003+A1:2011 standard.
There are eight primary defect categories, each image containing one major defect to ensure annotation clarity.
Privacy-sensitive information (e.g., text overlays) was automatically detected via EasyOCR (>95% accuracy) and blurred with OpenCV.
| Defect Class | EN 13508-2 Code | Number of Images | % of Total |
|---|
| Cracks / Breaks / Collapses | BAB, BAC | 4,654 | 34.65% |
| Intruding Sealing Material | BAI | 996 | 7.42% |
| Roots | BBA | 631 | 4.70% |
| Displaced Joint | BAJ | 1,389 | 10.34% |
| Pipe Surface Damage | BAF | 919 | 6.84% |
| Infiltration | BBF | 573 | 4.27% |
| Attached Deposits | BBB | 2,462 | 18.33% |
| Settled Deposits | BBC | 853 | 6.35% |
| No Defect | – | 1,014 | 7.55% |
| TOTAL | | 13,491 | 100% |
In total, the dataset includes 13,491 images.
Model Details
- Architectures: RT-DETR v1/v2, YOLO v8/v11/v12
- Framework: PyTorch → exported to ONNX
- Input: CCTV sewer inspection images
- Output: Predictions across 8 defect categories
Citation
If you use these models in your research, please cite our paper:
Mustafa Oğurlu, Bülent Bayram, Bahadır Kulavuz, Tolga Bakırman,
"Deep Learning for Automated Sewer Defect Detection: Benchmarking YOLO and RT-DETR on the Istanbul Dataset",
Applied Sciences, 2025.
DOI: [DOI will be here when available]
How to Use
1import onnxruntime as ort
2import numpy as np
3
4# Load model
5session = ort.InferenceSession("model.onnx")
6
7# Example: run inference (replace with your preprocessing)
8outputs = session.run(None, {"input": np.random.rand(1,3,640,640).astype(np.float32)})
9print(outputs)
Contact
For questions or collaboration requests, please contact:
mogurlu@gmail.com
