Official implementation of our paper.
Window to Wall Ratios (WWR) are key to assessing the performance of buildings, such as the energy consumption, daylight conditions, and ventilation 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. Therefore, to set up simulations a standard WWR is typically assumed for all buildings, or, at best, it's applied on a building-by-building basis based on initial knowledge regarding the building type and/or age. This paper presents a methodology to predict WWR of buildings from external street view images using semantic segmentation to detect windows and wall regions from images.
To generate the dataset we use a publicly available Window Detection dataset created using LabelMe, consisting of 2806 facade images.
We create two kinds of image pair datasets for the FCN Architecture and the Segformer architecture.
Original images are resized and normalized for the FCN architecture.
We obtain SegFormer model labels including exterior windows, by using a pre-trained SegFormer with a ResNet50dilated + PPMdeepsup encoder and decoder architecture to detect the building area in the images, and overlay window masks developed for the FCN architecture(above). We converted the output to greyscale masks encoding a single number per pixel, representing the encoded semantic label in the ADE20K semantic categories.
The dataset can be found here, and the code to prepare the Segformer dataset is provided here:
We train two models a Fully Convolutional Network (FCN) and a Segformer model, using the two datasets developed for these tasks outlined above.
The code to train the FCN can is provided here: 
The code to finetune the Segformer model is provided here: