The location and age of a real estate has a significant influence on its value. The goal of this project is to develop image analysis methods for the mining of age- and location-specific visual patterns from buildings as well as for their automated classification.

Real Estate Image Analysis (REIA) is a novel research initiative initiated by real estate experts and computer scientists from the Kufstein University of Applied Sciences and the St. Pölten University of Applied Sciences with the goal to improve the assessment of real estate by automated image analysis. The REIA initiative brings together two research disciplines that have so far hardly touched each other and bears a strong potential for novel and innovative real estate-related applications and services.

The theoretic foundation for real estate assessment is the hedonic pricing method which describes how the quantity and quality of object characteristics determine its price in a particular market. More formally, the hedonic price function takes the form: P_i = f(S_i, L_i, N_i), where P_i is the logarithm of the price or rent of house i, S_i represents structural housing characteristics, L_i location variables and N_i neighborhood characteristics. Traditional hedonic pricing is restricted to those characteristics which are available as quantifiable data, e.g. distances to schools, public transport, etc. REIA extends this model by additional - so far inaccessible - characteristics, which are derived by automated image analysis. To this end image data at different levels is leveraged. Structural housing characteristics (S_i) can be extracted from inside and outside views of buildings. Neighborhood (N_i) and location characteristics (L_i) in turn can be mined from 360° streetview and satellite images. The extracted characteristics represent a completely new set of parameters for real estate assessment and - beyond this - enable novel ways to index, search and retrieve real estate.

The project ImmoAge (FFG Bridge 855784) aims at automatically estimating the age and region of a building from an unconstrained photograph of its external view. We employ deep representation learning and transfer learning to discover visual patterns (e.g. building elements like windows, and doors) that are characteristic for certain building epochs (e.g. the 1960s or 1980s) and regions (e.g. Tyrol and Burgenland) and build classification and regression models from these representations.

Project partners


Despotivic, M., M. Sakeena, D. Koch, M. Döller, and M. Zeppelzauer, "Predicting Heating Energy Demand by Computer Vision", Computer Science - Research and Development, 09/2017.