Associative Trails

Classification of Building Information Model (BIM) Structures with Deep Learning (2018)

Notes on a 2018 paper by Francesco Lomio, Ricardo Farinha, Mauri Laasonen, and Heikki Huttunen

A comparison of different machine learning methods and their accuracy in categorising building use. The deep learning models came out on top, with the model based on ResNet50 performing particularly well. 240 images were used to train the models to sort into three categories.

Read the paper: IEEE, ResearchGate, PDF.

  • 240 images from BIM software (CAD isometrics)
  • 4 images form each of 60 BIM representations of buildings
  • 3 target categories - Apartment building, Industrial Building, Other
  • Random augmentation - horizontal flips, random rotations (not accounting for gravity?), horizontal and vertical shifts. This increased the learning data.
  • Analysed using 4 methods:
    • Feature extraction using Histogram of Oriented Gradients (HOG) followed by Support Vector Machine (SVM) classification
    • MobileNet - a pre-trained CNN. Last dense layer substituted for one with three categories. Width multiplier and resolution multiplier. Values of 0.25, 0.5, 0.75 & 1 assigned to width multiplier to validate each model.
    • Residual network (ResNet) - another pre-trained CNN. Specifically ResNet50. Last dense layer substituted for one with three categories.
    • Random structure CNN from scratch. Ended up with 3 convolutional layers, each with 16 feature maps and a kernel size of 2. No batch normalisation layer. Dropout rate of 0.3. Stochastic Gradient Descent, learning rate of 0.008130275. Each convolutional layer uses ReLU activation functions, with each layer downsampled by 2 with a max-pooling. Last convolutional layer flattens feature maps before the final 3-category dense layer.
  • Input images 224x224 to match inputs for the pre-trained CNNs
  • Results:
    Model Accuracy
    HOG + SVM 57.19% ± 1.18%
    ResNet50 97.92% ± 1.32%
    MobileNet α = 1 93.75% ± 2.94%
    MobileNet α = 0.75 95.42% ± 2.75%
    MobileNet α = 0.5 94.62% ± 3.35%
    MobileNet α = 0.25 91.70% ± 2.58%
    Custom CNN 89.60% ± 3.39%
 Jump to top