Underwater archaeological object detection using photogrammetric fusion

Abstract

The sea holds many secrets, yet the surveillance of underwater scenery still demands complex operations and sharp observation. Underwater conditions pose several challenges, including blurriness, degradation, and light distortion, all proving to be detrimental to detection performance. These challenging conditions have attracted researchers’ attention, leading to consistent improvements in detection accuracy. Building on these advancements, our work focuses on maritime archaeology sites, specifically detecting artefacts from the Tower Wreck in the Xlendi Underwater Archaeological Park, some even dating back to 300 BCE. This was done through the compilation of a multi-class dataset for underwater amphora detection, labelled by field experts. The final dataset covers 625m2 and consists of 864 images. Based on this dataset, various object detection architectures, including single-shot detectors (YOLO), transformer-based models (DETR), and two-stage detectors (Faster R-CNN), were evaluated. Transformer and two-stage models were found to underperform, possibly linked to their increased data requirements. On the other hand, YOLOv7- tiny achieved the best overall performance with a mAP50 of 86.14%. Further analysis was performed by using depth maps generated using the photogrammetric model, and saliency estimation techniques. Various fusion methods were then compared, which mixed these maps into the original imagery. This modification provided marginal improvements over base models, with YOLOv7-tiny mAP50 increasing to 86.34%. Finally, photogrammetric techniques were used to project 2D detections to 3D coordinates on the orthomodel. From these coordinates, several methods of processing were compared to best display these in an aggregated orthomosaic visualisation. These findings highlight the potential for bridging the gap between maritime archaeology and computer vision, paving the way for more efficient and accurate underwater archaeological surveys.