PSI - Issue 44

Agnese Natali et al. / Procedia Structural Integrity 44 (2023) 2020–2027 Agnese Natali et al./Structural Integrity Procedia 00 (2022) 000–000

2025

6

3.4. PAVIS-FABRE dataset To facilitate the development of a smart tool for recognizing different infrastructure elements, we create the PAVIS-FABRE dataset using 338 annotated images. The resolution of the images in this dataset ranges from 1280x720 to 4896x3672 pixels which were acquired using various cameras during visual inspections of existing structures performed by an expert team in the Fabre Consortium, which is an organization of research units which currently focus on the monitoring of the existing bridges. The dataset contains 6 annotation classes, viz. Armatura ossidata-corrosa (Corroded steel re-bars), Distacco del copriferro (Concrete cover detachment), Fusto (Column), Pila (Pier), Pulvino (Pulvinus) and Tracce di scolo (Drainage tracks). A total of 2744 regions are annotated in these images by 5 annotators using the CVAT annotation tool by Sekachev et al. (2019). A distribution of the annotations across various classes is shown in Fig. 7 and examples of images annotated with CVAT are shown in Fig. 8. We then present the images to our model with Xception feature extraction module trained on the CODEBRIM dataset. Qualitative results obtained on a few images from the PAVIS-FABRE dataset without retraining are shown in Fig. 9.

Fig. 6. Distribution of the annotations across various classes in the PAVIS-FABRE dataset.

Fig. 7. Annotated samples from the PAVIS-FABRE dataset. Images were annotated using the CVAT annotation tool Sekachev et al. (2019).

Fig. 8. Qualitative results obtained using our model trained with the CODEBRIM dataset on images from the PAVIS-FABRE dataset, without retraining.