Property archetype classifier
A DenseNet-201 that reads a photo of a UK house, from Google Street View or an upload, and dates its archetype from Georgian to contemporary. Served as a Flask web app and an AWS SageMaker endpoint inside Q-Bot's insulation survey system.
Problem
Q-Bot’s robots suit particular houses: the right age and construction, with a suspended timber floor and air vents. A house’s archetype, the era it was built in, is a strong first signal. Sending surveyors door to door to find out does not scale.
What I did
As the only person on the project, I curated a dataset from Google Street View images and fine-tuned a DenseNet-201 to classify a property into eight classes: Georgian (1714–1837), Victorian (1837–1870), Edwardian-Late-Victorian (1870–1910), Interwar (1918–1939), Postwar (1950–1980), Contemporary (1980+), Contemporary-Victorian (1980+), and Not-Visible/Greenery, for when a hedge is all the camera sees. Images are resized to 256 px with global contrast normalisation before the network. Alongside archetype, I worked on related attributes: wall type and texture, form, detachment and air vents.
I shipped it two ways. A Flask and Gunicorn web app in Docker, where you upload a photo and get the predicted archetype (the video above); a later version added confidence bars for every class. And an AWS SageMaker PyTorch endpoint that takes an image URL or raw bytes, plus a batch transform job for scoring large sets of addresses at once. I documented the SageMaker packaging and deployment for the team.
Result
The model was deployed as part of Q-Bot’s wider insulation survey system, screening properties from street imagery before anyone visits. Data, model and deployment were mine end to end.
figures
click to enlarge