Skip to content

Files

Latest commit

 

History

History
 
 

about

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

Project tasks

Gathering and pre-processing of data.

  • We have developed an original - and more importantly, automated - method of acquisition and preparation of training data.
  • Using publicly available resources, we have collected RGB photos for over 50,000 locations.
  • Using the Land and Buildings Register and some image transformations, a binary mask was prepared for each location, delineating the area of all registered buildings.

Development of a building segmentation model.

  • We conducted a series of experiments, during which we checked approx. 30 architectures, parameters and variants of deep neural network training.
  • We validated all models, which allowed us to choose the best - convolutional encoder-decoder (U-NET) model, with the F1 measure result over 0.78.

Preparation of the model API and the demo version of the system.

  • The final model was compressed using Tensorflow Lite, which allowed to reduce the amount of memory needed for prediction by 1/3.
  • A sample of the possibilities along with the visualization of the results is presented by a simple demo-application prepared with the use of Plotly and Dash libraries.

Importantly, only free tools and data sources were used in the process of creating the project.

Examples

Below, we present some examples of our model predictions - both those we are proud of and other that we still have to work on. ;)