Three different architectural blogs where chosen for this project: ArchDaily, Minimalissimo, SuckerPunchDaily. For each blog a specific scraping script extracting different data available from the specific post structure such as image links, post name, post text, year and author was constructed. This scraping step was done for all posts available on each of the selected websites – 53733 posts for ArchDaily, 2463 posts for Minimalissimo and 3171 posts for SuckerPunchDaily. For dimensionality-reduction each post image is encoded with a pretrained residual neural network – ResNet50 – trained on the imageNet dataset. This reduced representation is then used to project each image into 2d space using UMAP.
All the scraped data including the image links, the ResNet encoding and the 2d coordinates of the mapping is saved for each post. By choosing specific coordinates in the mapping space and evaluating distances to the mapped images, neighbourhoods can be extracted. These neighbourhoods show similar characteristics and often correspond to different concepts present in the images – indoor, outdoor, plans, winter, people, fashion, city. It is also possible to extract neighbours for a new image in one of the mappings by encoding it with the ResNet and then embedding it into the mapped space. By blending together 30 neighbour images of chosen coordinates or a mapped external image spectrum images of these localized concepts can be created.
