Window to Wall Ratio Detection using SegFormer

TL;DR

This work addresses the scarcity of accurate window-to-wall ratio (WWR) data by leveraging semantic segmentation on street-view façades to predict WWR, moving beyond common 40% heuristics. It introduces an augmented segmentation dataset that labels windows and façades, and evaluates FCN and SegFormer architectures to detect windows and estimate WWR, complemented by a four-point perspective correction method for frontal-view approximation. The FCN-ResNet-50 model achieves competitive IoU (~0.66) and transformer-based SegFormer variants show performance gains with larger models, while robustness and perspective corrections reveal challenges like glare and occlusions. Overall, the study demonstrates the feasibility of data-driven WWR estimation from exterior imagery and provides open-source code to advance architectural façade analytics.

Abstract

Window to Wall Ratios (WWR) are key to assessing the energy, daylight and ventilation performance of buildings. Studies have shown that window area has a large impact on building performance and simulation. However, data to set up these environmental models and simulations is typically not available. Instead, a standard 40% WWR is typically assumed for all buildings. This paper leverages existing computer vision window detection methods to predict WWR of buildings from external street view images using semantic segmentation, demonstrating the potential for adapting established computer vision technique in architectural applications

Figures (6)

• Figure 1: Dataset preparation: JPEG of building façade (left) and PNG with marked windows(right).
• Figure 2: Enhancing ADE20K labels by superimposing using SegFormer predicted labels (Left) onto our labeled windows (Right).
• Figure 3: Window pixels % vs Building Pixel %
• Figure 4: Prediction generated by FCN (left) and SegFormer (right) models on images in the test set
• Figure 5: FCN ResNet-50 predictions on varying levels of zoom and lighting conditions.