Abstract:
Despite the challenges in modern digital documentation, current research prioritizes computer-aided semantic segmentation
in underwater environments and temporal monitoring, particularly for the digital documentation of deep-sea sites. Using cutting-edge
technologies, exemplified by our automated archetype of archaeological sites (e.g., the Xlendi shipwreck), we present research on an
archaeological shipwreck known as Xlendi, located off the coast of Malta, aiming to facilitate digital model acquisition for professionals
and amateurs. This enhances archaeological insights and yields promising results across challenging sites, promoting virtual exploration,
awareness, and advocacy for underwater cultural heritage(UCH).
Indubitably, current 3D instance segmentation methods enhance archaeological site comprehension, but, they encounter challenges
such as computational complexity and labor-intensive annotation. This article addresses these issues by utilizing automated 2D object
detection extended to 3D through photogrammetry, minimizing human effort by focusing on ad-hoc 2D annotation methods seen in
previous research, and facilitating 3D segmentation through 2D 3D projection via photogrammetry.
intriguingly, the construction of this proposed model relies heavily on achieving precise 3D detection and identification. Its success
is contingent upon the performance of the 2D object detection and its projections in an end-to-end scene. In this study, we evaluate
the performance of YOLOv8 for object detection, focusing on underwater archaeological sites. Previous research using YOLOv4
reported an accuracy range of 78%-88% (mAP). Building on this, we assessed YOLOv8 using sensitivity, specificity, and mean average
precision (mAP), achieving mAP values ranging from 98.2% to 99.2%. Specifically, we measured mAP@0.50 and mAP@0.50:0.95 to
comprehensively evaluate model performance. Our findings demonstrate significant improvements over previous methods, highlighting
the efficacy of YOLOv8 in archaeological contexts. We have also included a future workflow to inspect further enhancements.
