Deep Learning for Building Detection for Completeness Check in Mottakskontroll – Master Thesis

This repository contains the code developed as part of a master's thesis at NTNU. The project investigates how deep learning can support the Norwegian Mapping Authority’s mottakskontroll (completeness check) for FKB-building data by detecting buildings in aerial imagery using instance segmentation models.

Project Structure

All the code produces in the master's thesis is in the master folder. The master folder have the following folders:

• mask_r_cnn: All code related to Mask R-CNN using Detectron2
• yolo: All code related to YOLOv8-seg using Ultralytics
• preprocessing: Scripts to tile TIFFs and generate training data
• scripts: Miscellaneous utilities (visualization, evaluation, etc.)

Models and Frameworks

The following models and tools are used in the project:

• YOLOv8-seg Ultralytics
• Mask R-CNN Detectron2
• SAHI – Sliced Aided Hyper Inference(SAHI) to improve predictions near tile edges

Installation

This project uses Python 3.12+. Dependencies are declared in the pyproject.toml file. Minimal installation example:

[project]
name = "master"
version = "0.1.0"
description = "Deep learning models for building detection in aerial images"
readme = "README.md"
requires-python = ">=3.12"
dependencies = [
    "torch>=2.6.0",
    "torchvision>=0.21.0",
]

To install dependencies, use a tool like pip or poetry:

pip install torch torchvision

⚠️ Additional installation steps are required for Ultralytics and Detectron2. Follow their official guides based on your OS and CUDA version.

Data Requirements

To run the full pipeline, you need:

• Georeferenced TIFF images
• FKB building data in polygon format stored as a GeoPackage (.gpkg)

These are processed into training tiles and annotations using the preprocessing/ scripts. Both COCO and YOLO format datasets are supported.

Usage

Preprocess the data Use the scripts in preprocessing/ to tile the input imagery and convert annotations to COCO or YOLO formats.

Train models

Run training scripts in yolo/ for YOLOv8-seg

Run training scripts in mask_r_cnn/ for Mask R-CNN (Detectron2)

Acknowledgements

This project was conducted as part of a master’s thesis in Engineering and ICT at NTNU, in collaboration with the Norwegian Mapping Authority (Kartverket) and the KartAI initiative.

Note

If you use this code or build upon it, please cite the frameworks used:

Ultralytics YOLOv8 Detectron2 SAHI

This repository is intended for academic and research use.