*3.2. Collection of Job Site Images for a Construction Project*

The multi-class classification in the image classification was selected for the study. Data sets were classified as rebar, worker, and machine. The deep learning model required massive amounts of information for training to improve its recognition accuracy, and the size of the data set was a critical factor for the experiment's success. Data came from three sources, as follows:


Four hundred sixty-one job site images were collected from the above sources (Figure 9). Feasible data were extracted from the images in the preliminary classification. The job site images collected were manually tagged for workers, machines, and rebar using the image tool provided in "LabelImg." In addition, movements were selected and tagged for classes.

**Figure 9.** Collection of construction site image files.

Two types of files were generated after tagging with LabelImg; one was the image files themselves, and the other was the XML files with image locations tagged. In Figure 10, for example, workers, rebars, and machines are tagged and given specific names in the image. Figure 11 provides an example of the contents of the XML file, including dimensions such as image coordinates. The single shot multibox detector (SSD) deep learning model was established and tested as all images were tagged.

**Figure 10.** LabelImg tagging of a job site photo.
