*3.1. Dataset Collection and Annotation*

Collection and annotation of images for building a concrete crack detector is very important stage because the accuracy of the trained model depends to large extent on the quality of the prepared dataset. For this research, 1058 images containing cracked concrete elements were collected from different laboratory experiments performed at Cracow University of Technology (CUT). Figure 5 shows selected images of concrete element containing cracks.

**Figure 5.** Example of images of concrete elements containing cracks.

The images were then manually annotated using a graphical image annotation tool LabelImg [37]. It works by defining a bounding box coordinates and the corresponding label. Figure 6 shows an example of image of concrete element containing several thin cracks.

**Figure 6.** Example of annotation of an image of concrete element containing several cracks.
