01
The paperBrown et al. · 2020
- Development of a CNN for the recognition and labelling of rooms within floor plan images of multi-residential apartments.
- Meeting a need to improve solar analysis compliance checking according to the local building code. Only living rooms and balconies need to be assessed, but these rooms were not defined in the BIM data.
- Input a floor plan, output an instance segmentation function where each room is classified and the region it occupies is masked.
- Using Action Design Research methodology - Investigate, act, observe.
- Contains a good review of other papers in the field
- Uses Python, TensorFlow, Keras, Visual Studio Code IDE.
- Data set of 454 floor plan images - CAD drawings, marketing images and hand-drawn floor plans. Compiling and classifying the training data proved to be the most extensive and time-consuming part of the project.
- Labels help in JSON files that defined the regions of rooms.
- 6 room classes used - Bedroom, Bathroom (including ensuite), Living room, Dining room, Kitchen, Balcony
- Labelled using VGG image annotator
- Instance segmentation - the classification of multiple class regions within a single image and produces a mask the defines the space the classes occupy within the image.
- Used 1000 epochs!
- Does not publish the results. Claims the results were “Satisfactory”, and “beginning to develop a fundamental understanding of design”.
- All training data came from one single architectural practice - did the CNN learn their “house style”?
We help professional services firms work out where AI genuinely helps, what has to be structured first, and what to leave alone.
Get in touch