Publikation
BLENDgänger: Generating Synthetic MBES Data for Underwater UXO Perception Tasks
Amos Smith; Nael Jaber; Leif Christensen; Martin Atzmueller
In: Jessica Daignault; Katie Skinner; Ed Verhamme (Hrsg.). OCEANS-2025 Great Lakes. OCEANS MTS/IEEE Conference (OCEANS-2025), September 29 - October 2, Chicago, Illinois, USA, Pages 1-9, IEEE, 2025.
Zusammenfassung
Unexploded ordnance (UXO) and discarded munitions in coastal waters pose serious environmental and safety risks. Effective explosive ordnance disposal (EOD) relies on accurate detection and characterization of UXO, often performed with multibeam echosounder (MBES) surveys. In practice, however, MBES data are corrupted by outliers stemming from sensor errors, environmental conditions, and acoustic interference. Machine-learning solutions demand large volumes of precisely labeled training data, but such labels are costly and timeconsuming to obtain. To overcome these challenges, we present BLENDgänger, a procedural data-generation framework built on the Blender platform. BLENDgänger synthesizes realistic MBES bathymetric point clouds with configurable noise profiles and ground-truth annotations, enabling the rapid assembly of largescale datasets for both semantic segmentation and 3D object detection of underwater ordnance. We demonstrate that models trained exclusively on BLENDgänger data achieve strong performance when evaluated on independent ex-situ MBES measurements. These results show that synthetic datasets can effectively bootstrap machine-learning workflows for UXO perception and inspection, reducing reliance on laborious manual annotation.