PSI - Issue 78

Simone Felicioni et al. / Procedia Structural Integrity 78 (2026) 1285–1292

1291

Fig. 5: Meta Quest Pro Headset employed in this study.

4.4. Results and discussion To assess the image-to-point cloud matching performance, we collect 54 images from different angles and lighting conditions of a structural crack. A qualitative result is shown in Fig. 6, while quantitative comparative results are reported in Table 2, demonstrating the superiority of the chosen technique compared to the baseline. More in detail, NetVLAD results to be completely unsuitable for this task, while Patch-NetVLAD benefits from the geometric check, which particularly enhances matching robustness in the scenarios under investigation where query and database images differ significantly in appearance. The reported metric, i.e., Recall@k, is commonly used in image retrieval tasks and measures how often the correct match is within the top-k results. In addition, Fig. 7 depicts a scene from the immersive virtual environment. Fig. 7a shows the keypoint associated with a structural damage, represented by a spherical interactive white marker. As mentioned, the position in the 3D space of the latter is known since the synthetic images in the database are retrieved by projecting the point cloud model into the 2D image space from strategical locations. Finally, Fig. 7b depicts a view collected by the MAV associated with that keypoint, with the possibility of scrolling through the acquired images for time-aware monitoring.

Table 2. Comparative results of image-to-point cloud matching. Recall@1 Recall@3

Recall@5

NetVLAD

0.00%

1.85%

5.56%

Patch-NetVLAD

77.78%

85.19%

88.89%

Fig. 6: Matching result between (a) the processed database image and (b) the query image.