A Comparative Analysis of Refraction-Aware SfM, Hierarchical Localization, and Gaussian Splatting for Underwater 3D Reconstruction

Abstract

Underwater 3D reconstruction requires handling both geometric distortion and degraded visual conditions. This paper compares three complementary methods: a refraction-aware Structure-from-Motion (RSfM) pipeline using Underwater Colmap (UW-Colmap), a deep learning-based Hierarchical Localization framework (HLOC), and a neural rendering approach using Gaussian Splatting (GS). The first applies nonlinear refraction correction via a modified Colmap pipeline to compensate for distortions introduced by flat-pane housings. It improves geometric consistency and reduces artifacts in tank and open-water captures but relies on accurate refractive modeling. HLOC enhances matching robustness in low-contrast and low-texture scenes using SuperPoint and SuperGlue. However, it introduces considerable noise, particularly with retrieval and exhaustive matching, resulting in degraded reconstruction accuracy without geometric correction. Gaussian Splatting provides real-time rendering of visually realistic scenes using 3D Gaussian primitives. While not designed for structural accuracy, it delivers high visual quality when supplied with calibrated poses. The paper’s core contribution is a controlled, side-by-side evaluation of these methods using a dual-environment dataset (air and underwater). By applying consistent evaluation metrics, geometry alignment, surface completeness, and visual consistency, we reveal the strengths and limitations of each approach. Results show that RSfM combined with GS provides the most reliable reconstruction and visualization pipeline. Deep learning methods are best applied at the feature level, followed by structured SfM for accurate geometry. This offers practical guidance for underwater photogrammetry and highlights the potential of hybrid reconstruction strategies.