Explicit vs Implicit Modelling of Refraction in Underwater Structure-from-Motion – A practical Guide

Abstract. This study presents a critical investigation, based on computer simulations, of refractive Structure-from-Motion (SfM) methods for underwater photogrammetry. The goal is to provide practical guidelines on when explicit refractive modeling is necessary and when it can be safely neglected for a given application. Simulations are run within the POSER framework (https://github.com/GEOSS-UNISS/POSER), which uses the physical based rendering engine of Blender to model the physical properties of water, including light refraction, scattering, and absorption phenomena. Simulated camera configurations range from a GoPro with a fisheye lens in a flat-port housing to a full-frame machine vision camera in a housing equipped with a dome port. Different offsets between the camera centre and the port geometry are also introduced to represent both standard and non-conventional housing setups, including misalignments. The study compares refractive and implicit (standard) modelling approaches, evaluating their respective advantages and limitations in terms of parameter estimability, the need for pre-calibration, and the availability of ground control points. Accuracy is assessed against ground-truth simulated data by analysing discrepancies in reconstructed 3D geometry and exterior orientation parameters.