Underwater archaeological object detection through bidirectional photogrammetric fusion

Abstract

Advancements in 3D reconstruction techniques have dramatically reduced the requirements to obtain accurate 3D models. However, the digitized scenes still require hours of manual labor to analyze and document, calling for improved automated interpretation tools. This study presents a framework for the bidirectional integration between object detection and photogrammetric modeling, as applied to an original multi-class object detection dataset captured during the Tower Project in Xlendi, Gozo. The photogrammetric surface is used to render depth maps, which are encoded into the red channel forming a process called RDMix. This process is based on the lack of red light observed in deep-sea sites, which offers additional bandwidth. By making more efficient use of preexisting resources, RDMix is also detector agnostic, being an enrichment tool suitable for many applications. After detection, 2D predictions are then projected back onto the model, enabling individual object geotagging and aggregated detection on the orthomosaic. From the experiments conducted, RDMix provides consistent marginal gains in detection accuracy, whilst the geotagging and projection processes enrich the tangible improvements of automated detection.