architectural-style-recognition
architectural-style-recognition copied to clipboard
dumitrux
ArchitecturalStyle Recognition is a image classifier of 25 different architectural styles, using deep learning.
ArchitecturalStyle Recognition
ArchitecturalStyle Recognition is a image classifier of 25 different architectural styles, using deep learning.
Made with fastai library. Source code on src. It can be runned in Google Colab
Table of content
- Dataset
- Training
- Results
- Demonstration
- Run the app locally
- License
- Links
Dataset
The dataset contains 10113 images from 25 architectural styles. It is a mixed between images scraped from Google Images and the dataset from the paper "Architectural Style Classification using Multinomial Latent Logistic Regression" (ECCV2014), made by Zhe Xu.
Number of images for each style
- Achaemenid architecture: 392
- American craftsman style: 364
- American Foursquare architecture: 362
- Ancient Egyptian architecture: 406
- Art Deco architecture: 566
- Art Nouveau architecture: 615
- Baroque architecture: 456
- Bauhaus architecture: 315
- Beaux-Arts architecture: 424
- Byzantine architecture: 313
- Chicago school architecture: 278
- Colonial architecture: 480
- Deconstructivism: 335
- Edwardian architecture: 280
- Georgian architecture: 381
- Gothic architecture: 331
- Greek Revival architecture: 523
- International style: 417
- Novelty architecture: 382
- Palladian architecture: 343
- Postmodern architecture: 322
- Queen Anne architecture: 720
- Romanesque architecture: 301
- Russian Revival architecture: 352
- Tudor Revival architecture: 455
The original dataset was made by Zhe Xu. According to the paper, the best accuracy they could get in 2014 was nearly 70% accuracy. Paper "Architectural Style Classification using Multinomial Latent Logistic Regression" (ECCV2014)
Data augmentation
The extended dataset is used to test the model with "real world" images, in this case google images, whicha are not taken in perfect conditions.
To make the original dataset bigger and get better results, data augmentation is used. Which is achived creating new images, by transforming (horizontal flip, zoom, light, rotate, etc.) the images.
In this case we go from 4979 to 9588 images.
Training
Training vs Test sets
The dataset is splitted in two:
- training set (80%)
- test set (20%)
Results
We got a 95% of accuracy on this dataset
Demonstration
More examples
Greek Revival:
Byzantine:
Gothic:
Bauhaus:
Deployment
The app folder contains the application ready to deploy on Render.
Run the app locally
The app can be runned locally.
First of all, you have to check all the Requirements, then run this command in your terminal:
python app/server.py serve
Or in python3:
python3 app/server.py serve
Note: it has been only tested in Linux.
License
MIT License
Links
Author
Danci, Marian Dumitru (@dumitrux)
dumitrux