Construction-sites-detection project

This repo contains code and dataset for training and testing ml model which implements instance segmentation of construction sites.

Architectures:

Yolo-v8 and Yolo-v11: https://github.com/ultralytics/ultralytics

CSDS: Construction Site Detection and Segmentation

The Construction Site Detection and Segmentation(CSDS) is a tool consisting of 16 CSDS models and of large-scale dataset of construction site satellite imagery with detailed polygon annotations(5 stages and footings).
It contains both the raw source data (images and XML annotations) and preprocessed training-ready splits in YOLO format.

Pretrained Models

Pretrained models are available at:
👉 issai/CSDS_models

Dataset

The dataset are available at:
👉 datasets/issai/CSDS_models To request access, please fill out the form. Access will be provided once it has been manually.

Dataset Structure

• All images and annotations are provided in ZIP archives for efficient storage and download.
• The raw/ folder contains original images and XML annotations.
• The preprocessed/ folder contains processed input images (600px and 1200px) with YOLO-style train/test/val splits.

Raw Data

raw/
├── images/                     # Original construction site images
└── annotations/
├── AOD/                     # XML annotations for *All Objects Dataset*
└── FVOD/                    # XML annotations for *Fully Visible Objects Dataset*

Preprocessed Data

preprocessed/
├── AOD/
│   ├── 600/                     # Images with input resolution of 600px (YOLO format)
│   │   ├── train/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   ├── val/
│   │   │   ├── images/
│   │   │   └── labels/
│   │   └── test/
│   │       ├── images/
│   │       └── labels/
│   └── 1200/                    # Images with input resolution of 1200px (YOLO format)
│       └── (train/test/val structure as above)
│
└── FVOD/
├── 600/
│   └── (train/test/val with images + labels)
└── 1200/
└── (train/test/val with images + labels)

• AOD/ → Preprocessed dataset corresponding to "all objects" annotations.
• FVOD/ → Preprocessed dataset corresponding to "fully visible objects" annotations.
• Each size folder (600/, 1200/) contains YOLO-ready train/, val/, and test/ splits with images/ and labels/ directories.

If you use the dataset/source code/pre-trained models in your research, please cite our work:

@dataset{issai_csds_2025,
  title        = {CSDS: AI-Based Construction Site Detection and Segmentation tool for Satellite Images},
  author       = {Bissarinova, U. and Awan, H. H. and Olagunju, S. O. and Bolatkhanov, I. and Turekhassim, A. and Varol, H. A. and Karaca, F.},
  year         = {2025},
  publisher    = {Institute of Smart Systems and Artificial Intelligence and Department of Civil Engineering, Nazarbayev University},
  howpublished = {\url{https://doi.org/10.48333/0PJD-BP65}}
}