Zenodo is currently experiencing slowness and intermittent outages due to heavy automated traffic from bots and AI crawlers. We are aware of the problem, and our team is focused on stabilizing the service. Thank you for your patience.
Automatic assessments of building plans are uncommon in the early design stages, especially when schematic sketches are in raster format. Existing design evaluation tools, such as fire code reviewers, which are typically used in the late design stage, primarily evaluate vector format images that contain complete building information. These tools use conditional shape-embedding techniques to analyze the vector images. However, there are limitations to identifying and evaluating drawings through vector-shape relationships. Our research aimed to develop tools that can automatically assess schematic sketches in raster format to overcome the limitations of existing tools. We integrated a conditional shape-embedding tool, named Shape Machine, to assess vector images, with machine learning techniques, namely a Generative Adversarial Network (GAN), to assess raster sketches. This integration enables the evaluation of fire evacuation sketches in the early stages of the design process, thereby improving design efficiency and reducing costs. Moreover, in the future, this integration could allow the evaluation of designs in multiple image formats.
Notes
We used Shape Machine (Economou et al. 2021), a technology based on shape grammar, to translate the fire codes into code statements and review the five key elements in building plans in vector format.
To build our dataset, the text in the fire code was translated into ten checking functions that help recognize vector image data in the generated plans and review their compliance with different fire code criteria (see Figure 2). We developed a labeling rule that used different colors to mark areas with different plan errors (see Figure 3). Colors with RGB values ranging from 0 to 255 were used to differentiate the labels as much as possible. Therefore, we used ten combinations of RGB values to label the ten types of errors corresponding to the different fire code criteria. If any element in the room did not meet the code requirements, the room was colored accordingly; for example, Room 1 in Figure 3 was labeled red (R:255 G:0 B:0) when breaking Code 4; similarly, four tones of green (R:0 G:63 B:0, R:0 G:63 B:0, R:0 G:63 B:0, and R:0 G:63 B:0) were used to label width errors when breaking Code 5, setting these drawing layers always on the top of the others.
We used Shape Machine (Economou et al. 2021) to translate each code in Table 2 into the checking functions listed in Table 3 for the recognition process. As depicted in Figure 4, Checking Functions 1.1 and 1.2 were used to examine whether the distance between two doors of a room or two exits was less than 5 m. Checking Function 2 was employed to verify if there were at least two doors when the room area was greater than 75 m2. Checking Function 3 was utilized to check whether the shortest distance between doors and exits, as calculated through the Grasshopper Shortest Walk plug-in, was less than 22 m. Checking Function 5 evaluated whether the distance between points and doors in a room was less than 22 m. Specifically, Checking Functions 5.1, 5.2, 5.3, and 5.4 were utilized to review the minimum widths of doors, exits, corridors, and stairways, respectively. Finally, Checking Function 6 determined if the distance between the doors and exits of the first floor was less than 15 m. To complete all the checking functions, we utilized Shape Machine to recognize the vector-embedded shapes of the five elements and substitute them with raster colors. Subsequently, the reviewed elements were automatically color-coded in raster format to build a machine-learning image dataset.
To access the full dataset, Please email tuhan_arch@outlook.com.
