Benchmarking 3D Reconstruction for
Under-Ice Robotic Perception
in Arctic Environments

Klára Orbán*1, Ashwin Nedungadi*2, Ana-Ștefania Kogălniceanu1, Bianca-Marcela Sava1, Bastian Cieply2, Raul-Alexandru Simon1, Kuderna-Iulian Bența1, Stefan Lüdtke2
1Babeș-Bolyai University, Cluj-Napoca, Romania   2University of Rostock, Germany
* Equal Contribution
Paper Dataset

Abstract

As the Arctic transitions toward a seasonally ice-free state, capturing the complex morphology of submerged sea ice is crucial for advancing models of ocean-ice-atmosphere heat exchange and monitoring under-ice habitats.

Currently, sub-surface ice mapping is predominantly achieved through acoustic sensors such as multibeam sonar. While sonar provides robust large-scale geometry, it lacks the spectral and high-resolution textural information required to record fine morphological features — a critical limitation in under-ice environments, where sparse visual features and degraded sensing conditions challenge the capture of detailed structural and optical surface properties.

We present a benchmark for vision-based under-ice 3D reconstruction using real-world data from the MOSAiC expedition and synthetic data generated in an Unreal Engine-based underwater robotics simulator. The benchmark evaluates Structure-From-Motion (SfM) and monocular depth estimation, with quantitative comparison enabled through cross-modal registration to upward-looking multibeam sonar.

1. A benchmark for evaluating vision-based 3D reconstruction methods against spatially aligned under-ice multibeam sonar data.

2. A cross-modal alignment procedure that defines a shared sensing region for quantitative comparison between optical and acoustic reconstructions.

3. A simulation pipeline in HoloOcean based on sonar-derived ice geometry for generating controlled multimodal under-ice datasets.

Ice 3D Reconstruction

Interactive 3D Model
Drag to orbit · Scroll to zoom

Interactive reconstruction — Under-ice geometry from DA3 applied to ROV imagery collected during the MOSAiC expedition. Drag to orbit, scroll to zoom.

Methods

We evaluate two complementary paradigms for vision-based 3D reconstruction against multibeam sonar ground truth from the MOSAiC expedition, using a shared cross-modal alignment procedure across real and simulated environments.

🧊

Structure-from-Motion (SfM)

Agisoft Metashape pipeline with ROV navigation priors for camera alignment, dense point cloud generation, and ICP-based cross-modal registration to the sonar reference.

🤖

Monocular Depth — DA3

Depth Anything V3 estimates per-frame depth maps and camera parameters from single images, enabling 3D reconstruction without explicit multi-view geometry.

🔊

Acoustic Ground Truth

Imagenex DT101 multibeam sonar at 240 kHz — 480 beams, 120° swath. Provides a physically grounded geometric reference unaffected by lighting or surface texture.

⚓

Cross-Modal Alignment

Manual coarse alignment followed by Iterative Closest Point (ICP) refinement. Evaluation restricted to the jointly observed region to avoid penalizing uncovered areas.

Area 1 — Cross-Modal Alignment

Cross-modal alignment Area 1

Fig. 1 — Cross-modal alignment in Area 1: composite overlay showing four-point registration between multibeam sonar geometry and coincident optical imagery.

DA3 Depth Estimation

DA3 depth comparison
DA3 reconstruction

Fig. 2 — (Top) Original ice image alongside its DA3 depth map. (Bottom) Example 3D reconstruction from four enhanced images using estimated camera poses.

Experimental Results

Results reveal a strong domain dependence. SfM excels on real-world MOSAiC imagery while DA3 performs better under controlled simulation conditions — reflecting fundamentally different sensitivities to lighting, feature texture, and optical scattering.

Metric SfM (Metashape) DA3 (Monocular)
RMSE (m)0.55251.2103
MAE (m)0.45080.8569
ARE0.01070.0204
W-RMSE* (m)0.57610.9705
Chamfer (m)1.70363.0413
Hausdorff (m)9.087113.4949
DA-Chamfer (m)1.69842.5935

* W-RMSE weighted by point cloud density for a representative global error estimate.

Metric SfM (Metashape) DA3 (Monocular)
RMSE (m)10.21490.3176
MAE (m)7.85510.2414
ARE0.31450.0097
W-RMSE (m)10.38630.2917
Chamfer (m)10.11270.7839
Hausdorff (m)79.55024.0488
DA-Chamfer (m)10.16100.7563

In simulation, DA3 benefits from controlled lighting and consistent surface appearance.

SfM – Sonar Alignment (CloudCompare)

CloudCompare C2C distance visualization

Fig. 3 — CloudCompare visualization of the alignment between the SfM reconstruction and the sonar reference. The sonar slice is shown in white; the SfM point cloud is colored by cloud-to-cloud (C2C) distance. Blue/green indicate good alignment; yellow/red indicate larger deviations.

HoloOcean Simulation Pipeline

The HoloOcean framework mitigates the inherent noise and gaps present within field-gathered MOSAiC data. Using Unreal Engine's hardware-efficient rendering, we deployed an eight-motor ROV platform in a controlled environment with customized camera and sonar configurations.

An ice mesh was imported from real sonar survey data, with custom materials approximating the optical conditions at the ice-water interface. Water column parameters were calibrated to simulate subsurface light scattering with dense volumetric fog and non-uniform light attenuation. The simulated ROV carries a synchronized RGB camera and multibeam sonar matched to the MOSAiC hardware geometry and sampling rate.

Simulated Environment vs. DA3 Reconstruction

HoloOcean vs DA3 point cloud

Fig. 4 — Comparison of the simulated environment (left: HoloOcean underwater viewport) and the resulting DA3 reconstruction (right: 3D point cloud from the simulated camera trajectory).

Depth Map Enhancement & Difference Heatmap

Depth heatmap

Fig. 5 — Per-pixel absolute difference (normalized) between original and enhanced grayscale depth maps. Preprocessing can refine local structure alignment but cannot fully overcome the sim-to-real gap.

Dataset — MOSAiC Expedition

The primary dataset was collected during the Multidisciplinary Drifting Observatory for the Study of Arctic Climate (MOSAiC) expedition, providing both under-ice topography and image data acquired with a multibeam sonar system and a high-definition camera mounted on a remotely operated vehicle (ROV).

We prioritized survey PS122/3_39-20, conducted on May 5, 2020, which provided the most visually consistent and geometrically complete overlap between acoustic and optical sensor streams.

Acoustic System

The Imagenex DT101 multibeam sonar operates at 240 kHz, emitting 480 beams simultaneously with a 120° across-track swath and 3° along-track beam width. The ROV was typically operated at 20 m below the ice for standard surveys (10 m for high-resolution passes). Standard processing yielded ~0.5 m grid resolution; high-resolution products achieved ~0.1 m lateral resolution.

Optical Processing

Videos required preprocessing to remove navigation overlays and watermarks. Upward-looking ROV photos were selected as the primary reconstruction input. Image enhancement (sharpening and haze reduction) was applied as a domain-alignment step for DA3, which is primarily trained on terrestrial imagery.

Work in progress

Future Work

Our results highlight two complementary directions for future research: expanding the diversity of synthetic training environments, and closing the gap between offline benchmarking and real-time autonomous operation.

We are actively developing an extended synthetic pipeline in HoloOcean and plan to release a new open dataset later this year. This involves generative mesh pipelines (Meshy, Pixal3D) to produce diverse ice morphologies on demand — from smooth congelation ice to heavily fractured pressure ridges — broadening training scenarios without requiring additional field expeditions.

Drone-based RGB and LiDAR surveys of the ice surface will be co-registered with underwater sonar to build cohesive full-thickness ice models, simultaneously capturing freeboard and draft for volume estimation and radiative transfer modelling. From a geometric perspective, the above-water surface can act as a guiding prior for camera pose estimation in regions where optical methods beneath the ice fail due to turbidity or featureless surfaces.

Synthetic multimodal data will be used to fine-tune vision foundation models such as DA3 against the photometric scattering and texture-sparse conditions inherent to under-ice environments, directly addressing the sim-to-real gap observed in this benchmark. Ultimately, we aim to translate this into a hybrid acoustic-visual SLAM pipeline, wherein sonar constrains global drift and optical flow refines local trajectory estimates, targeting computationally efficient deployment on embedded ROV hardware at survey framerates.

Synthetic Pipeline Preview

Meshy-based composite iceberg model rendered in HoloOcean with irregular above-water geometry and submerged draft.

HoloOcean upward-looking still image of subsurface part of an ice structure.

Drone/co-registration view in simulated environment with underwater ice.

BibTeX

@article{orban2025underice,
  author    = {Orbán, Klára and Nedungadi, Ashwin and
               Kogălniceanu, Ana-Ștefania and Sava, Bianca-Marcela and
               Cieply, Bastian and Simon, Raul-Alexandru and
               Bența, Kuderna-Iulian and Lüdtke, Stefan},
  title     = {Benchmarking 3D Reconstruction for Under-Ice
               Robotic Perception in Arctic Environments},
  year      = {2025},
}

Acknowledgments

The authors express thanks to the MOSAiC expedition team for facilitating the collection of the under-ice data used in this research, and to the Alfred Wegener Institute (AWI) for support in the data collection process. Special thanks go to Florian Zimmer for the dataset and technical support throughout the research process.

Data used in this manuscript was produced as part of the international Multidisciplinary drifting Observatory for the Study of the Arctic Climate (MOSAiC) with the tag MOSAiC20192020 and the Project ID AWI_PS122_00.