SFS-A68: a dataset for the segmentation of space functions in apartment buildings

TL;DR

This paper introduces SFS-A68, a new dataset of images from apartment building models, enabling deep learning-based segmentation of space functions to improve building analysis tasks.

Contribution

The paper provides the first dataset for deep learning-based space function segmentation in apartment buildings and evaluates initial models demonstrating its feasibility.

Findings

Deep learning models can effectively segment space functions.

The dataset supports development of automated space classification.

Transfer learning improves segmentation accuracy.

Abstract

Analyzing building models for usable area, building safety, or energy analysis requires function classification data of spaces and related objects. Automated space function classification is desirable to reduce input model preparation effort and errors. Existing space function classifiers use space feature vectors or space connectivity graphs as input. The application of deep learning (DL) image segmentation methods to space function classification has not been studied. As an initial step towards addressing this gap, we present a dataset, SFS-A68, that consists of input and ground truth images generated from 68 digital 3D models of space layouts of apartment buildings. The dataset is suitable for developing DL models for space function segmentation. We use the dataset to train and evaluate an experimental space function segmentation network based on transfer learning and training from scratch. Test results confirm the applicability of DL image segmentation for space function classification. The code and the dataset of the experiments are publicly available online (https://github.com/A2Amir/SFS-A68).