ZEMCH 2019 International Conference Proceedings April.2020 | Page 369
Table 1. Dataset Composition
WBS Level 1
Structural Wall
Subtotal Amount
Architectural Wall
Subtotal Amount
Total Amount
WBS Level 2
Concrete Wall
Masonry Wall
Tile Wall
Drywall
Quantity of Data
3,947
3,062
7,009
3,248
3,185
6,433
13,442
Status
beginning/in progress
beginning/in progress
beginning/in progress
beginning/in progress
2.1.2. Convolutional Neural Network (CNN) Model Development
There has been rapid development in computer vision‐based shape recognition technologies, and
thus, a great variety of developed models are available. In this paper, the model to be studied is a
Convolutional Neural Network (CNN) model, which is efficient when the number of behaviors or objects
to be identified in the images is singular or small amount.
The CNN model applied in this study is called Standard CNN or SCNN. It is a fundamental model
and has model layers including convolutional layer, Rectified Linear Unit (ReLU), Max‐Pooling, and
output layers. Table 2. shows the other types of CNN models and their features.
The models created by the authors were written in Python version 3.6 with Anaconda prompt
environment using the Jupiter notebook, and the specific computer language is captured in Figure 2. For
an experiment process, after training the CNN model with labeled dataset images, a new picture of
construction site is input. Then the corresponding WBS level 1 and 2 are displayed as output values, which
are recorded as a format of html.
Table 2. List of Convolutional Neural Networks
Name
Standard CNN
(S‐CNN)
VGGNet 16
Faster RCNN
YOLO
Feature
‐Basic Convolutional Neural Network
‐Convolutional Layer, ReLU, Max‐Pooling Layer, Output Layer [4, 5]
‐Deeper architecture with small convolutional filters [6, 7]
‐14 Layers of Convolutional Layer, ReLU, Max‐Pooling Layer
‐Faster learning rate compared to the number of layers
‐Region Proposal Network sharing fill‐image convolutional feature [8, 9]
‐CNN(VGG) feature map, Intermediate Layer, Output Layer
‐Model can search multiple features in one images, yet requires fine‐tuned
CNN
‐Real‐time object detection with a single CNN [10, 11]
‐24 Convolutional Layers, two fully Connected Layers
Computer Vision and 4D BIM Applied Construction Project Progress Monitoring
358