Benchmarking Neural Rendering Approaches for 3D Reconstruction of
Underwater Environments
Salvatore Mario Carota
1,∗ a
, Alessandro Privitera
1,∗ b
, Daniele Di Mauro
2 c
,
Antonino Furnari
1,2 d
, Giovanni Maria Farinella
1,2 e
and Francesco Ragusa
1,2 f
1
Department of Mathematics and Computer Science, University of Catania, Viale Andrea Doria, 6, Catania, Italy
2
Next Vision s.r.l., Viale Andrea Doria, 6, Catania, Italy
Keywords:
Underwater 3D Reconstruction, Neural Rendering, 3D Gaussian Splatting.
Abstract:
We tackle the problem of 3D reconstruction of underwater scenarios using neural rendering techniques. We
propose a benchmark adopting the SeaThru-NeRF dataset, performing a systematic analysis by comparing
several established methods based on NERF and 3D Gaussian Splatting through a series of experiments. The
results were evaluated both quantitatively, using various 2D and 3D metrics, and qualitatively, through a user
survey assessing the fidelity of the reconstructed images. This serves to provide critical insight into how to
select the optimal techniques for 3D reconstruction of underwater scenarios. The results indicate that, in the
context of this application, among the algorithms tested, NeRF-based methods performed better in both mesh
generation and novel view synthesis than the 3D Gaussian Splatting based methods.
1 INTRODUCTION
3D reconstruction is a classic computer vision task
that has become ubiquitous across various scien-
tific fields, including archaeological inspections (De
Reu et al., 2014), biological studies (Correia and
Brito, 2023; Irschick et al., 2022), and architectural
projects (M
¨
unster et al., 2024; Cui et al., 2024).
An area of particular interest, due to its diverse ap-
plications ranging from biological assessment to ar-
chaeological discovery, is underwater 3D reconstruc-
tion, which poses unique challenges due to several
critical differences compared to reconstructing non-
underwater scenes. Image captured in underwater en-
vironments differ significantly because of the pres-
ence of water, which alters the behavior of light (Li
et al., 2019; Islam et al., 2020; Zhang and Johnson-
Roberson, 2023; Hou et al., 2020). These differ-
ences include variations in lighting, optical distor-
a
https://orcid.org/0009-0008-6431-9156
b
https://orcid.org/0009-0001-3507-8233
c
https://orcid.org/0000-0002-4286-2050
d
https://orcid.org/0000-0001-6911-0302
e
https://orcid.org/0000-0002-6034-0432
f
https://orcid.org/0000-0002-6368-1910
∗
These authors share first authorship.
tions, and limited visibility. Together, these factors
create a complex set of challenges for accurate 3D
reconstruction (Akkaynak and Treibitz, 2019). Over-
coming these challenges is highly beneficial for many
fields. In underwater heritage conservation, 3D re-
construction enables the inspection of artifacts and
structures without risking damage or compromising
their integrity (Memet, 2008; Perez-Alvaro, 2023).
This technology not only aids in protecting cultural
assets but also allows for their presentation to a wider
audience, such as in virtual museums. Additionally,
underwater environmental sciences can benefit from
advancements in 3D reconstruction technologies to
monitor coral reef health by detecting changes over
time. Detailed 3D models enable marine biologists to
study complex habitats, providing deeper insights into
ecological interactions (Zhang et al., 2023; Adam-
czak et al., 2019; Kaandorp, 1993). 3D reconstruc-
tions of underwater environments can also be utilized
in video games, movies, and virtual and augmented
reality applications to enhance user experiences. A
great example of its usage in the field of culture is a
project called ”First Life”
1
, which enables the visitor
to become a virtual voyager, traveling through sub-
1
https://www.nhm.ac.uk/discover/news/2015/june/dive-
back-in-time-with-david-attenborough-s-first-life.html
766
Carota, S. M., Privitera, A., Di Mauro, D., Furnari, A., Farinella, G. M. and Ragusa, F.
Benchmarking Neural Rendering Approaches for 3D Reconstruction of Underwater Environments.
DOI: 10.5220/0013381200003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages
766-773
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

Acquired Underwater Images 3D Reconstruction
NERF
3D Gaussian Splatting
Figure 1: From a set of acquired images in underwater environments, the task is to reconstructs the 3D model of the environ-
ment.
merged landscapes and creating for themselves a new
perspective on the history of life on Earth. Recently,
several solutions have dealt with the problem of 3D
environment reconstruction, but a big gap remains in
underwater environments reconstruction. Photogram-
metry (Sch
¨
onberger and Frahm, 2016), Neural Radi-
ance Fields (NeRF) (Levy et al., 2023), and 3D Gaus-
sian splatting (3DGS) (Kerbl et al., 2023) techniques
have played an important role in enhancing and intro-
ducing new methods for reconstruction of 3D models.
In this work, we present a benchmark for
3D Reconstruction of underwater environments (see
Figure 1) using the state-of-the-art SeaThru-NeRF
dataset introduced in (Levy et al., 2023). The bench-
mark compares models based on NeRF and 3DGS,
at the same time we test how underwater enhancing
techniques perform as a preprocessing step in the con-
text of neural rendering. The results were evaluated
both quantitatively, using a variety of evaluation met-
rics at both the rendering and mesh generation levels,
and qualitatively, through a user survey. We found
out that NeRF-based models are slightly better suited
for the task compared to 3DGS-based methods, which
nonetheless remain highly promising.
The contributions of this work are: 1) We con-
ducted a systematic analysis of NeRF-based and
3DGS-based methods for underwater environment re-
construction, providing insights into the performance
of the tested methods; 2) We analyzed how enhance-
ment techniques can improve the reconstruction of
3D models; 3) We quantitatively evaluated both new
view synthesis task and 3D mesh reconstruction; 4)
We conducted a qualitative study on the accuracy of
reconstructed 3D models through questionnaires ad-
ministered to a total of 40 subjects.
2 RELATED WORK
Our work builds on prior research in underwater
datasets, underwater image enhancement, 3D recon-
struction, and neural rendering, which will be briefly
described in the following sections.
Underwater Datasets. In literature several dataset
were proposed (Li et al., 2019; Islam et al., 2020;
Zhang and Johnson-Roberson, 2023; Hou et al., 2020;
Akkaynak and Treibitz, 2019), some of them are real,
i.e. capturing real scenes, others are synthetic, i.e. im-
ages are crafted in some way to solve a particular task.
They are used for various purposes, i.e. enhancement,
3d reconstruction, robotics. Among them: UIEB (Li
et al., 2019) consisting of 950 real-world underwa-
ter images with different natural and artificial light-
ing conditions. UFO-120 (Islam et al., 2020) contains
1,500 paired samples splitted in training and valida-
tion sets and 120 paired samples for benchmark eval-
uation. Each shot is provided with a high-resolution
ground truth version, its distorted low-resolution ver-
sion, and a saliency map mask. BNU (Zhang and
Johnson-Roberson, 2023) includes images captured
in a 1.3m-deep tank and in Lake Erie. The JPEG
images were post-processed and camera poses were
calculated using COLMAP. SUID (Hou et al., 2020),
is a synthetic dataset produced by applying special ef-
fects that simulate underwater conditions in terrestrial
images. SeaThru and the following work from the
same research group SeaThru-NERF (Akkaynak and
Treibitz, 2019; Levy et al., 2023) contains underwa-
ter scenes captured in three different sea with a total
of 29, 20 and 18 images respectively. We chose this
dataset for training the models due to its diverse range
of scenarios and number of images.
Underwater Image Enhancement and Restoration.
Underwater image enhancement is the task to reduce
or remove water effects on images recorded under-
water. WaterGAN (Li et al., 2017) is a color cor-
rection model based on Generative Adversarial Net-
works (GAN). The generator estimates the attenua-
tion, backscatter, and camera characteristics of un-
derwater images. The model is trained with both
underwater and non-underwater images in order to
create synthetic underwater images. In (Cho et al.,
2020) they used GANs for image correction and en-
hancement through the image-translation technique.
Benchmarking Neural Rendering Approaches for 3D Reconstruction of Underwater Environments
767

Figure 2: Images belonging to the Seathru dataset adopted for the proposed benchmark. The dataset contains three different
underwater scenes: Red Sea (left), Caribbean Sea (center) and Pacific Ocean (right).
The model is trained with underwater images in or-
der to capture textures and details of underwater im-
ages, the losses used are reconstruction loss, lapla-
cian loss and perception loss. Furthermore, Semi-
UIR (Huang et al., 2023) is a semi-supervised un-
derwater image restoration framework based on the
mean-teacher model, designed to incorporate unla-
beled data into network training. The student model
learns from labeled data, while the teacher model
guides the training process on unlabeled images by
generating reliable “pseudo-labels.” Experimental re-
sults on both full-reference and no-reference under-
water benchmarks show significant improvements in
both quantitative and qualitative performance over
SOTA methods. We used this method for the enhance-
ment preprocessing step.
Multi-view surface reconstruction is the process
that creates a 3D surface starting from a set of im-
ages taken from different angles, exploiting points
correspondences between images to estimate the
shape of an object. Several works adopted volumet-
ric grid methods for reconstructing multi-view sur-
faces (Boent and Pula, 1999; Kutulakos and Seitz,
2000; Laurentini, 1994; Szeliski, 1993; Seitz and
Dyer, 1999). Other works focused on the tech-
niques of cloud point-based techniques (Furukawa
and Ponce, 2009; Galliani et al., 2015; Schoenberger
et al., 2016; Tola et al., 2012). Its influence has, how-
ever been even more extensive among dense recon-
struction methods: the Poisson surface reconstruc-
tion algorithm (Kazhdan et al., 2006a) along with its
screened version (Kazhdan and Hoppe, 2013). Ma-
chine learning methods adopting deep learning for en-
hancing multi-view surface reconstruction techniques
have also recently been explored (Chen et al., 2019;
Huang et al., 2018; Yao et al., 2018).
Neural Radiance Fields (NeRF) is a pioneering ap-
proach that was initiated by (Mildenhall et al., 2020)
The idea is to use a neural network to implicitly model
a scene from a set of images annotated with a pose.
The model thus learns the behavior of light and the
geometry of a scene enabling the use of such model
to generate novel views. Several variants have been
developed to extend the possibilities of NeRF (Bar-
ron et al., 2021; Alex et al., 2021; Kai et al., 2020;
Sun et al., 2022).
Neural surface reconstruction is a technique that in-
volves using neural networks to learn complex sur-
faces of 3D objects or scenes in a continuous and
highly detailed manner. Various methods have been
proposed for this task, utilizing volumetric grid-based
methods for scene reconstruction (Niemeyer et al.,
2020; Oechsle et al., 2021). The authors of (Lior
et al., 2020) uses Signed Distance Functions (SDF)
to implicitly model surfaces by defining them as the
zero level set of SDF. Similar NeRF-based methods
have further been extended to surface reconstruction,
works such as (Wang et al., 2021; Lior et al., 2021;
Darmon et al., 2022; Fu et al., 2022; Yue et al.,
2022) have moved the frontiers in extending the orig-
inal NeRF framework towards high-fidelity surface
modeling. There are also point cloud-based tech-
niques (Fu et al., 2022; Zhang et al., 2022), which
accomplished good reconstructions taking in input
sparser data points.
3D Gaussian Splatting (3DGS) was proposed
in (Kerbl et al., 2023). The method is a change in
perspective compared to NeRF, 3D scenes are repre-
sented in an explicit way: the scene is modelled as a
collection of 3D Gaussian functions distributed in the
space then they are ”splatted” in 2D in order to match
the set of images, together with the corresponding
cameras calibrated by Structure from Motion, taken
in input. The core of the approach is the optimiza-
tion step, where a dense set of 3D Gaussians accu-
rately representing the scene is created. In addition
to positions and covariance, it also optimizes Spheri-
cal Harmonics coefficients representing color of each
Gaussian to correctly capture the view-dependent ap-
pearance of the scene. The optimization of these
parameters is interleaved with steps that adaptevely
control the density of the Gaussians to better repre-
sent the scene. The optimization takes full advantage
of standard GPU-accelerated frameworks and adds
custom CUDA kernels, following recent best prac-
tices (Alex et al., 2021; Sun et al., 2022). The pro-
jection method implements a tile-based rasterizer for
Gaussian splats inspired by recent software rasteriza-
tion approaches (Lassner and Zollhofer, 2021). The
rasterization pipeline is fully differentiable, and given
the projection to 2D can rasterize anisotropic splats
similar to previous 2D splatting methods (Kopanas
et al., 2021). As for NeRF, various method are im-
proving over the initial, such as Splatfacto-W (Xu
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
768

Figure 3: Original image sample (left) vs. its enhanced version (right).
et al., 2024), which improves results in presence of
unconstrained photo collections. Its key contributions
include latent appearance modeling, efficient tran-
sient object handling, and precise background mod-
eling.
Underwater 3D Reconstruction. Two major tech-
niques are used for 3D reconstruction: image-
based (Levy et al., 2023; Weidner et al., 2017;
Jordt et al., 2016) and laser-scanner-based (Bartolini
et al., 2005). Image-based methods are highly cost-
effective, while laser-scanner-based methods require
expensive equipment and typically take a long time
to acquire data. In particular SeaThru-NeRF (Levy
et al., 2023) is a NeRF-based method specifically de-
signed for underwater scenes, with the unique capa-
bility of modelling the solid objects present in the
scene and the medium.
3 BENCHMARK
Dataset. We perform our benchmark using the
dataset presented in (Levy et al., 2023). It con-
tains underwater scenes captured in three different
seas (see Figure 2): the Red Sea (Eilat, Israel),
the Caribbean Sea (Curacao), and the Pacific Ocean
(Panama), with a total of 29, 20, and 18 images re-
spectively. The images were acquired as RAW im-
ages using a Nikon D850 SLR camera in a Nauticam
underwater housing with a dome port to avoid refrac-
tions. The images were resized to an average size of
900 × 1400 and white-balanced with a 0.5% clipping
per channel to remove extremely noisy pixels. Fi-
nally, COLMAP (Sch
¨
onberger and Frahm, 2016) was
used to extract the camera poses.
Task. We consider the problem of generate a 3D
model from a set of underwater images, where the
same scene is captured from different viewpoints.
Models. We trained the following models: Neuralan-
gelo (Li et al., 2023), Seathru-NeRF (Levy et al.,
2023), Splatfacto (Kerbl et al., 2023), and Splatfacto-
W (Xu et al., 2024). We also performed a COLMAP
dense reconstruction which is our baseline. We trans-
formed the outputs of each model into 3D meshes
using different methods. For Neuralangelo, we used
the Marching Cubes algorithm (Lorensen and Cline,
1987). SeaThru-NeRF, on the other hand, was pro-
cessed using Poisson Surface Reconstruction (Kazh-
dan et al., 2006b). For Splatfacto and Splatfacto-
W, we employed TSDF (Truncated Signed Distance
Function) implemented in dn-splatter (Turkulainen
et al., 2024). Lastly, COLMAP dense reconstruc-
tion was also processed using Poisson Surface Re-
construction (Kazhdan et al., 2006b). We also evalu-
ated the opportunity to do image enhancement before
using Neural Rendering Models (see Figure 3). The
enhanced dataset was created by applying the SEMI-
UIR algorithm, trained on SUID dataset, to the im-
ages of SEATHRU-NeRF. Results that consider en-
hanced images are indicated with +enh.
We finally imported these meshes into Blender
2
to generate renders (see Figure 4) using camera paths
exported directly from Nerfstudio
3
.
Quantitative Evaluation. To evaluate the qual-
ity of rendered images, we adopted various metrics:
MUSIQ (Ke et al., 2021) gives a score of the per-
ceived quality of an image-very similar to human
judgments as well as UCIQE (Yang and Sowmya,
2015) and UIQM (Panetta et al., 2016) are specif-
ically designed for quality assessment in underwa-
ter images. We also evaluated the mesh generation
quality using the Hausdorff distance (Cignoni et al.,
1998). In particular, we used the implemented ver-
sion in MeshLab
4
. This allowed us to make a quan-
titative comparison of the meshes resulting from the
various algorithms with respect to a dense reconstruc-
tion from COLMAP. We computed the metric bidirec-
tionally, obtaining a symmetric version by taking the
maximum value. Differently from the other metrics,
which evaluate 2D results, this distance evaluates the
accuracy of the reconstruction of the 3D geometry of
the environment.
Qualitative Evaluation. To evaluate the quality of
2
https://www.blender.org/
3
https://docs.nerf.studio/
4
https://www.meshlab.net/
Benchmarking Neural Rendering Approaches for 3D Reconstruction of Underwater Environments
769

Figure 4: The image shows 3D model renders (Curac¸ao scene) created by four different algorithms and rendered in Blender
(the green border indicates the ground truth).
renderings obtained by the different reconstruction
models, we designed a survey that has been admin-
istered to 40 people. Participants were asked to rate
each 3D reconstruction on a scale from 1 to 5, where
a score of 1 indicates the lowest quality and 5 repre-
sents the best quality.
4 RESULTS
Table 1 shows the average results derived from cal-
culating the different 2D metrics across all images
rendered from a set of sampled viewpoints along a
camera path, evaluated for each scene and model.
For the metric MUSIQ: NeRF-Based methods lead
the scoreboard with 2 best results (Neuralangelo-
Panama 61.513, SeaThru-NeRF-Readsea 65.612) and
one second best. For the metric UCIQE. NeRF-
Based methods lead the scoreboard with 3 best results
(Neuralangelo-Panama 3.461, Neuralangelo-Readsea
5.769, Neuralangelo-Enh-Curac¸ao 4.62) and 3 second
best. Finally for UIQM: gaussian splatting method
Splatfacto-W+enh is the best performer. We can
state that 2 metrics over 3 show an advantage of
NeRF-based metrics. Table 2 compares the rendered
outputs to a reference image calculating the PSNR
among them. The results show that Seathru-NeRF
and COLMAP are the leading models in terms of
rendering accuracy, Splatfacto-W also performs well
(see Mean column: 15.8329 vs. 15.8203 vs 15.4112).
Here we have a second proof that a NeRF based
method is favorable to gaussian splatting ones. Fi-
nally 3D Mesh distances using Hausdorff distance are
reported in Table 3. Seathru-NeRF achieves excellent
performance because it can produce a good recon-
struction, thanks to the ability to distinguish between
medium and objects. SeaThru-NeRF is closely fol-
lowed by Splatfacto-W+enh (0.0482 vs 0.0715), here,
we can observe how enhancement plays a crucial role.
All the quantitative measures show a clear advan-
tage by NeRF on gaussian splatting, in particular with
Figure 5: Qualitative evaluation results.
SeaThru-NeRF leading the scoreboard for 3D mesh
generation and PSNR results. For the qualitative as-
sessment, from the survey results we noticed that
SeaThru-NeRF consistently outperformed other mod-
els, emerging as the best reconstruction model in two
out of the three scenes. Splatfacto-W + enh. was iden-
tified as the second-best model overall, also achieving
high ratings across most scenes. This ranking high-
lights the relative strengths of these two models in
producing high-quality renderings for diverse scenes.
Figure 4 shows some examples of the reconstructions
provided to users for evaluation.
These tables collectively provide insight into the
relative strengths of each model. In our view, Seathru-
NeRF and Splatfacto-W achieve superior results pri-
marily due to their specialized features. Seathru-
NeRF is specifically designed for underwater envi-
ronments, making it particularly effective in handling
underwater visual challenges. Splatfacto-W, benefits
from advanced background modeling and effective
handling of transient objects, both of which enhance
its rendering accuracy and adaptability.
Qualitative results in Figure 5 are coherent with
quantitative ones: SeaThru-NeRF consistently out-
performed other models, emerging as the best re-
construction model in two out of the three scenes.
Splatfacto-W + Enh was identified as the second-best
model overall, also achieving high ratings across most
scenes. Some caveats: the survey was not random-
ized, thus some bias could be present.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
770

Table 1: MUSIQ (Ke et al., 2021), UCIQE (Yang and Sowmya, 2015) and UIQM (Panetta et al., 2016) evaluation of new view
syntesis, best results in bold, second best underlined. Each value represents the average of multiple rendered image values.
Curac¸ao Panama Redsea
Model MUSIQ UCIQE UIQM MUSIQ UCIQE UIQM MUSIQ UCIQE UIQM
Neuralangelo 46.336 2.411 2.138 61.513 3.461 2.475 49.186 5.769 1.351
Neuralangelo+enh. 50.824 4.62 1.95 58.524 2.232 2.868 56.885 4.367 1.911
SeaThru-NeRF 57.412 4.427 2.145 55.432 0.852 1.927 65.612 0.654 2.021
Splatfacto 59.274 1.896 1.703 49.674 1.391 1.617 61.045 1.38 1.560
Splatfacto-w 61.470 2.836 2.478 55.694 1.123 2.073 64.632 1.069 1.983
Splatfacto +enh. 59.742 1.733 1.843 56.055 1.443 1.623 59.860 1.404 1.567
Splatfacto-w+enh. 64.042 1.940 2.564 58.066 1.185 2.260 63.105 1.099 2.052
Colmap-Poisson 45.350 1.132 1.712 51.647 0.983 1.838 65.289 0.802 1.410
Table 2: Average PSNR, rendered image is compared to
original image or to the original preprocessed when en-
hancement is in place, best results in bold, second best un-
derlined.
Method Curac¸ao Panama Redsea Mean
Neuralangelo 10.6242 11.3654 9.4266 10.4721
Neuralangelo+enh. 10.0941 10.1605 9.7282 9.9943
SeaThru 16.1436 16.7314 14.6238 15.8329
Splactfacto 17.3554 15.1125 10.7050 14.3910
Splactfacto-W 16.4640 16.1713 13.5984 15.4112
Splactfacto+enh. 10.5768 10.2170 9.8972 10.2303
Splactfacto-W+enh. 11.1514 11.3191 11.6711 11.3805
Colmap-Poisson 18.4726 16.4224 12.566 15.8203
Table 3: Mean Normalized Haussdorf distance (Cignoni
et al., 1998) between COLMAP reconstruction and
NERF/3DGS based reconstruction, best results in bold, sec-
ond best underlined.
Method Curac¸ao Panama Redsea Mean
Neuralangelo 0.1538 0.0969 0.1387 0.1298
Neuralangelo+enh. 0.1706 0.0819 0.1379 0.1301
SeaThru-NeRF 0.0287 0.0652 0.0508 0.0482
Splatfacto 0.0906 0.0466 0.1394 0.0922
Splatfacto-W 0.0753 0.0539 0.0919 0.0737
Splatfacto+enh. 0.1347 0.0818 0.1203 0.1123
Splatfacto-W+enh. 0.0809 0.0517 0.0818 0.0715
Colmap-Poisson 0 0 0 0
5 CONCLUSION
In this work, we presented a benchmark to perform
3D reconstruction of underwater scenes using Neural
Rendering techniques. Quantitative analysis shows
promising results, Neural Rendering model are on par
with SfM only when they take care of modeling the
medium (SeaThru-NeRF) or when they take care to
model different camera settings (Splatfacto-W). In fu-
ture work we will focus on improving 3DGS with
medium modelling similar to Seathru-NeRF.
ACKNOWLEDGEMENTS
This research has been supported by Next Vi-
sion s.r.l.
5
and by the project Neural Rendering &
Edge AI Platform for 4D synthetic Twins generation
during Underwater Navigation & Exploration (NEP-
TUNE) PNRR MUR Project CUP J53D23020140005
COR 18115262 - Spoke 3 Robotics and AI for Socio-
economic Empowerment (RAISE).
REFERENCES
Adamczak, S. K., Pabst, A., McLellan, W. A., and Thorne,
L. H. (2019). Using 3d models to improve estimates of
marine mammal size and external morphology. Fron-
tiers in Marine Science, 6.
Akkaynak, D. and Treibitz, T. (2019). Sea-thru: A method
for removing water from underwater images. In Pro-
ceedings of the IEEE/CVF conference on computer vi-
sion and pattern recognition, pages 1682–1691.
Alex, Y., Fridovish-Keil, S., Matthew, T., Qinhong, C., Ben-
jamin, R., and Angjoo, K. (2021). Plenoxels: Radi-
ance fields without neural networks. arXiv preprint
arXiv:2112.05131.
Barron, J. T., Mildenhall, B., Tancik, M., Hedman, P.,
Martin-Brualla, R., and Srinivasan, P. P. (2021). Mip-
nerf: A multiscale representation for anti-aliasing
neural radiance fields. In Proceedings of the
IEEE/CVF International Conference on Computer Vi-
sion, pages 5855–5864.
Bartolini, L., De Dominicis, L., Ferri de Collibus, M., For-
netti, G., Guarneri, M., Paglia, E., Poggi, C., and
Ricci, R. (2005). Underwater three-dimensional imag-
ing with an amplitude-modulated laser radar at a 405
nm wavelength. Applied optics, 44(33):7130–7135.
Boent, J. S. and Pula, P. (1999). Probabilistic voxelized vol-
ume reconstruction. In Proceedings of International
Conference on Computer Vision (ICCV), volume 2.
Chen, R., Han, S., Xu, J., and Su, H. (2019). Point-based
multi-view stereo network. In Proceedings of the
IEEE/CVF International Conference on Computer Vi-
sion, pages 1538–1547.
5
https://www.nextvisionlab.it/
Benchmarking Neural Rendering Approaches for 3D Reconstruction of Underwater Environments
771

Cho, Y., Jang, H., Malav, R., Pandey, G., and Kim, A.
(2020). Underwater image dehazing via unpaired
image-to-image translation. International Journal of
Control, Automation and Systems, 18:605–614.
Cignoni, P., Rocchini, C., and Scopigno, R. (1998). Metro:
measuring error on simplified surfaces. In Computer
Graphics Forum, volume 17, pages 167–174. Black-
well Publishers.
Correia, H. A. and Brito, J. H. (2023). 3d reconstruction
of human bodies from single-view and multi-view im-
ages: A systematic review. Computer Methods and
Programs in Biomedicine, 239:107620.
Cui, D., Wang, W., Hu, W., Peng, J., Zhao, Y., Zhang,
Y., and Wang, J. (2024). 3d reconstruction of build-
ing structures incorporating neural radiation fields and
geometric constraints. Automation in Construction,
165:105517.
Darmon, F., Basc
´
ole, B., Devaux, J.-C., Souhila, P., and
Aubry, M. (2022). Improving neural implicit surfaces
geometry with patch warping. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition, pages 6260–6269.
De Reu, J., De Smedt, P., Herremans, D., Van Meirvenne,
M., Laloo, P., and De Clercq, W. (2014). On introduc-
ing an image-based 3d reconstruction method in ar-
chaeological excavation practice. Journal of Archae-
ological Science, 41:251–262.
Fu, Q., Sun, Q., Yew, T.-W., and Tiao, W. (2022). Geo-
neus: Geometry-consistent neural implicit surfaces
learning for multi-view reconstruction. arXiv preprint
arXiv:2205.15848.
Furukawa, Y. and Ponce, J. (2009). Accurate, dense, and ro-
bust multiview stereopsis. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 32(8):1362–
1376.
Galliani, S., Lasinger, K., and Schindler, K. (2015). Mas-
sively parallel multiview stereopsis by surface normal
diffusion. In Proceedings of the IEEE International
Conference on Computer Vision, pages 873–881.
Hou, G., Zhao, X., Pan, Z., Yang, H., Tan, L., and Li, J.
(2020). Benchmarking underwater image enhance-
ment and restoration, and beyond. IEEE Access,
8:122078–122091.
Huang, P.-H., Kopf, J., Ahuja, N., Bleyer, M., Lenz, J.,
and Xu, J.-B. (2018). Deepmvs: Learning multi-view
stereopsis. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pages
2821–2830.
Huang, S., Wang, K., Liu, H., Chen, J., and Li, Y. (2023).
Contrastive semi-supervised learning for underwater
image restoration via reliable bank. In Proceedings
of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 18145–18155.
Irschick, D. J., Christiansen, F., Hammerschlag, N., Martin,
J., Madsen, P. T., Wyneken, J., Brooks, A., Gleiss, A.,
Fossette, S., Siler, C., Gamble, T., Fish, F., Siebert,
U., Patel, J., Xu, Z., Kalogerakis, E., Medina, J.,
Mukherji, A., Mandica, M., Zotos, S., Detwiler, J.,
Perot, B., and Lauder, G. (2022). 3d visualization pro-
cesses for recreating and studying organismal form.
iScience, 25(9):104867.
Islam, M. J., Luo, P., and Sattar, J. (2020). Simulta-
neous Enhancement and Super-Resolution of Under-
water Imagery for Improved Visual Perception. In
Robotics: Science and Systems (RSS), Corvalis, Ore-
gon, USA.
Jordt, A., K
¨
oser, K., and Koch, R. (2016). Refractive 3d
reconstruction on underwater images. Methods in
Oceanography, 15-16:90–113. Computer Vision in
Oceanography.
Kaandorp, J. A. (1993). 2d and 3d modelling of ma-
rine sessile organisms. In Crilly, A. J., Earnshaw,
R. A., and Jones, H., editors, Applications of Fractals
and Chaos, pages 41–61, Berlin, Heidelberg. Springer
Berlin Heidelberg.
Kai, Z., Gernot, R., Noah, S., and Vladlen, K. (2020).
Nerf++: Analyzing and improving neural radiance
fields. arXiv preprint arXiv:2010.07492.
Kazhdan, M., Bolitho, M., and Hoppe, H. (2006a). Poisson
surface reconstruction. In Proceedings of the Fourth
Eurographics Symposium on Geometry Processing,
pages 61–70.
Kazhdan, M., Bolitho, M., and Hoppe, H. (2006b). Poisson
surface reconstruction. In Proceedings of the Fourth
Eurographics Symposium on Geometry Processing,
SGP ’06, page 61–70, Goslar, DEU. Eurographics As-
sociation.
Kazhdan, M. and Hoppe, H. (2013). Screened poisson sur-
face reconstruction. ACM Transactions on Graphics
(ToG), 32(3):1–13.
Ke, J., Wang, Q., Wang, Y., Milanfar, P., and Yang, F.
(2021). Musiq: Multi-scale image quality transformer.
In Proceedings of the IEEE/CVF International Con-
ference on Computer Vision, pages 5148–5157.
Kerbl, B., Kopanas, G., Leimk
¨
uhler, T., and Drettakis, G.
(2023). 3d gaussian splatting for real-time radiance
field rendering. ACM Trans. Graph., 42(4):139–1.
Kopanas, G., Philip, J., Leimk
¨
uhler, T., and Drettakis, G.
(2021). Point-based neural rendering with per-view
optimization. In Computer Graphics Forum, vol-
ume 40, pages 29–43. Wiley Online Library.
Kutulakos, K. N. and Seitz, S. M. (2000). A theory of shape
by space carving. International Journal of Computer
Vision, 38(3):199–218.
Lassner, C. and Zollhofer, M. (2021). Pulsar: Efficient
sphere-based neural rendering. In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition, pages 1440–1449.
Laurentini, A. (1994). The visual hull concept for
silhouette-based image understanding. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence,
16(2):150–162.
Levy, D., Peleg, A., Pearl, N., Rosenbaum, D., Akkaynak,
D., Korman, S., and Treibitz, T. (2023). Seathru-nerf:
Neural radiance fields in scattering media. In Proceed-
ings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition, pages 56–65.
Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., and
Tao, D. (2019). An underwater image enhancement
benchmark dataset and beyond. IEEE Transactions
on Image Processing, 29:4376–4389.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
772

Li, J., Skinner, K. A., Eustice, R. M., and Johnson-
Roberson, M. (2017). Watergan: Unsupervised gener-
ative network to enable real-time color correction of
monocular underwater images. IEEE Robotics and
Automation letters, 3(1):387–394.
Li, Z., M
¨
uller, T., Evans, A., Taylor, R. H., Unberath, M.,
Liu, M.-Y., and Lin, C.-H. (2023). Neuralangelo:
High-fidelity neural surface reconstruction. In IEEE
Conference on Computer Vision and Pattern Recogni-
tion (CVPR).
Lior, Y., Yoni, K., Dror, M., Matan, A., Meirav, R., and
Yaron, L. (2020). Multiview neural surface recon-
struction by disentangling geometry and appearance.
Advances in Neural Information Processing Systems,
33:2492–2503.
Lior, Y., Yoni, K., Dror, M., Matan, A., Meirav, R., and
Yaron, L. (2021). Volume rendering of neural implicit
surfaces. Advances in Neural Information Processing
Systems, 34:4805–4815.
Lorensen, W. E. and Cline, H. E. (1987). Marching cubes:
A high resolution 3d surface construction algorithm.
SIGGRAPH Comput. Graph., 21(4):163–169.
Memet, J.-B. (2008). Conservation of underwater cultural
heritage: characteristics and new technologies. Mu-
seum International, 60(4):42–49.
Mildenhall, B., Srinivasan, P. P., Tancik, M., Barron, J. T.,
Ramamoorthi, R., and Ng, R. (2020). Nerf: Repre-
senting scenes as neural radiance fields for view syn-
thesis. In ECCV.
M
¨
unster, S., Apollonio, F. I., Bl
¨
umel, I., Fallavollita, F.,
Foschi, R., Grellert, M., Ioannides, M., Jahn, P. H.,
Kurdiovsky, R., Kuroczy
´
nski, P., Lutteroth, J.-E.,
Messemer, H., and Schelbert, G. (2024). Handbook
of digital 3d reconstruction of historical architecture.
page 204.
Niemeyer, M., Mescheder, L., Oechsle, M., and Geiger, A.
(2020). Differentiable volumetric rendering: Learning
implicit 3d representations without 3d supervision. In
Proceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition (CVPR), pages
3504–3515.
Oechsle, M., Peng, S., and Geiger, A. (2021). Unisurf:
Unifying neural implicit surfaces and radiance fields
for multi-view reconstruction. In Proceedings of the
IEEE/CVF International Conference on Computer Vi-
sion, pages 5589–5598.
Panetta, K., Gao, C., and Agaian, S. (2016). Human-visual-
system-inspired underwater image quality measures.
IEEE Journal of Oceanic Engineering, 41(3):541–
551.
Perez-Alvaro, E. (2023). Underwater cultural heritage and
the sustainable development goals. Blue Papers, 2(2).
Schoenberger, J. L., Zheng, E., Frahm, J.-M., and Pollefeys,
M. (2016). Pixelwise view selection for unstructured
multi-view stereo. In Proceedings of the European
Conference on Computer Vision, pages 501–518.
Sch
¨
onberger, J. L. and Frahm, J.-M. (2016). Structure-
from-motion revisited. In Conference on Computer
Vision and Pattern Recognition (CVPR).
Seitz, S. M. and Dyer, C. R. (1999). Photorealistic scene re-
construction by voxel coloring. International Journal
of Computer Vision, 35(2):151–173.
Sun, C., Sun, M., and Chen, H.-T. (2022). Direct voxel
grid optimization: Super-fast convergence for radi-
ance fields reconstruction. In Proceedings of the
IEEE/CVF conference on computer vision and pattern
recognition, pages 5459–5469.
Szeliski, R. (1993). Rapid octree construction from image
sequences. CVGIP: Image Understanding, 58(1):23–
32.
Tola, E., Strecha, C., and Fua, P. (2012). Efficient large-
scale multi-view stereo for ultra high-resolution image
sets. Machine Vision and Applications, 23(5):903–
920.
Turkulainen, M., Ren, X., Melekhov, I., Seiskari, O., Rahtu,
E., and Kannala, J. (2024). Dn-splatter: Depth and
normal priors for gaussian splatting and meshing.
Wang, Y., Skorokhodov, I., Theobalt, P., and Wonka, P.
(2021). Hf-neus: Improved surface reconstruction us-
ing high-frequency details. Advances in Neural Infor-
mation Processing Systems, 34:19220–19230.
Weidner, N., Rahman, S., Li, A. Q., and Rekleitis, I. (2017).
Underwater cave mapping using stereo vision. In 2017
IEEE International Conference on Robotics and Au-
tomation (ICRA), pages 5709–5715. IEEE.
Xu, C., Kerr, J., and Kanazawa, A. (2024). Splatfacto-
w: A nerfstudio implementation of gaussian splatting
for unconstrained photo collections. arXiv preprint
arXiv:2407.12306.
Yang, M. and Sowmya, A. (2015). An underwater color
image quality evaluation metric. IEEE Transactions
on Image Processing, 24(12):6062–6071.
Yao, Y., Luo, Z., Li, S., Fang, T., and Quan, L. (2018).
Mvsnet: Depth inference for unstructured multi-view
stereo. In Proceedings of the European Conference on
Computer Vision (ECCV), pages 767–783.
Yue, Z., Peng, S., Niemeyer, M., Sattler, T., and Geiger,
A. (2022). Exploring monocular geometric cues for
neural implicit surface reconstruction. arXiv preprint
arXiv:2206.00665.
Zhang, C., Zhou, H., Christiansen, F., Hao, Y., Wang, K.,
Kou, Z., Chen, R., Min, J., Davis, R., and Wang, D.
(2023). Marine mammal morphometrics: 3d model-
ing and estimation validation. Frontiers in Marine Sci-
ence, 10.
Zhang, J., Shao, Y., Li, T., Fang, D. M., Tsian, Y., and
Quan, L. (2022). Critical regularizations for neural
surface reconstruction in the wild. In Proceedings of
the IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 6270–6279.
Zhang, T. and Johnson-Roberson, M. (2023). Beyond nerf
underwater: Learning neural reflectance fields for true
color correction of marine imagery. IEEE Robotics
and Automation Letters, 8(10):6467–6474.
Benchmarking Neural Rendering Approaches for 3D Reconstruction of Underwater Environments
773