Multiview Space Function Classification in Apartment Buildings Using Image Deep-Learning Semantic Segmentation
Amir Ziaee, Georg Suter · Journal of Computing in Civil Engineering · 2025
Image deep learning semantic segmentation (IDLSS) models have significant application potential for automating classification tasks in building design and engineering. Applying IDLSS to space function classification is promising due to the rich context features in images of space layouts. Previous studies are limited to space function segmentation (SFS) of single-view rather than multiview space layout images. The multiview approach may improve predictive performance by better capturing the shapes of spaces and space elements and reducing view bias. Moreover, existing SFS models have been applied to layouts of single functional units, such as apartments, rather than entire floors, which is more desirable but more challenging. Our study addresses both limitations. We created and compared 54 SFS models based on state-of-the-art IDLSS using multiview images of apartment building floors. For this purpose, we created a data set, which includes multiview input and ground truth images generated from 68 3D digital space layout models of apartment buildings. The top-performing SFS model had a mean IoU of 65.2% using transfer learning. We evaluated several effects on its predictive performance. Excluding space elements, such as doors or sanitary elements, from input multiview space layout images resulted in an 8.6% decrease in mean IoU, indicating that space elements provide essential context features. Using single-view instead of 16-view images for training and testing resulted in an 18% decrease in mean IoU, suggesting a superior predictive performance of the multiview versus the single-view approach. We detected view bias, which decreases as the number of views increases. We developed an algorithm to extract space-level predictions from pixel-level predictions in SFS output images. We applied it to 16-view test images predicted by the top-performing SFS model. Results show that the predictive performance was generally higher for larger spaces and spaces with context features. It was lower for smaller, windowless spaces.