BlueCoral3D: Underwater Benchmark Dataset

• BlueCoral3D is a real-world underwater dataset featuring shallow reef scenes under dynamic caustic illumination, captured with calibrated 360° image sweeps.
• The dataset comprises around 500 high-resolution images per scene with precise camera calibration, enabling robust multi-view consistency analysis.
• Evaluation focuses on novel-view synthesis using photometric metrics (PSNR, SSIM, LPIPS), highlighting relative reconstruction quality rather than absolute 3D accuracy.

Searching arXiv for the specified papers to ground the article in the cited sources. BlueCoral3D is a real-world underwater dataset introduced in the context of “From Restoration to Reconstruction: Rethinking 3D Gaussian Splatting for Underwater Scenes” (Huang et al., 22 Sep 2025). It is intended for view-synthesis and relative reconstruction evaluation in shallow-water reef environments under dynamic surface-caustic illumination. The dataset consists of two natural scenes with approximately 500 high-resolution images per scene, full camera calibration for every frame, and a held-out-view evaluation protocol based on photometric metrics rather than absolute 3D geometric error. The available description presents BlueCoral3D primarily as a challenging benchmark for underwater novel-view synthesis and reconstruction under naturally varying illumination, while also highlighting the absence of explicit 3D ground-truth geometry, semantic labels, and detailed physical metadata.

1. Definition and scope

BlueCoral3D consists of 2 real-world underwater scenes. Both are natural shallow-water reef environments featuring diverse coral formations and rocks; no shipwrecks or synthetic tanks are included (Huang et al., 22 Sep 2025). Both scenes are in clear, shallow coastal waters subject to dynamic surface-caustic illumination, that is, sunlight refracted by waves. No turbid or deep-water scenes are included.

This scope places BlueCoral3D within a relatively specific operating regime: clear, shallow, caustic-lit reef imagery rather than a general-purpose underwater benchmark. A plausible implication is that the dataset is especially useful for studying failure modes of reconstruction systems under illumination instability, but not for evaluating robustness across the full range of underwater optical conditions.

2. Data acquisition and camera calibration

Each scene contains approximately 500 high-resolution images, captured in a full 360° sweep around the scene (Huang et al., 22 Sep 2025). The image resolution is 1920 × 1080. Although the source paper mentions video sequences, no explicit frame-rate is given; in practice the sweep was captured at typical underwater video rates of approximately 24–30 Hz.

Intrinsic parameters, including focal length and principal point, and full extrinsics, specifically 6 DOF poses, are provided for every frame, calibrated via standard multi-view calibration (Huang et al., 22 Sep 2025). This makes the dataset directly usable for reconstruction pipelines that assume known camera parameters and pose trajectories.

The combination of dense viewpoint coverage and per-frame calibration indicates that BlueCoral3D is organized around multi-view consistency rather than sparse surveying. This suggests a benchmark design aligned with NeRF-style and 3DGS-style rendering systems, where pose quality and view overlap materially affect both optimization stability and held-out rendering performance.

3. Ground truth, annotation, and metadata

The paper does not provide any explicit 3D ground-truth geometry, such as laser scanning or structured light (Huang et al., 22 Sep 2025). No depth maps, meshes, or dense point clouds are included. BlueCoral3D is intended for view-synthesis and relative reconstruction evaluation only, using held-out images as “ground truth” for photometric metrics.

No per-pixel semantic labels are supplied, and illumination and water properties such as scattering or absorption coefficients are not measured or released (Huang et al., 22 Sep 2025). No auxiliary physical metadata, including temperature and salinity, accompany the dataset.

These omissions materially constrain the kinds of claims that can be made from experiments on BlueCoral3D. In particular, no absolute 3D geometric error, for example point-to-point RMSE, can be computed because ground-truth depth is unavailable. Similarly, the lack of semantic labels and physical metadata precludes direct evaluation of joint semantic-photometric reconstruction or physics-informed rendering.

4. Evaluation protocol and metrics

Each scene is split by assigning every 100th image to a test set; the remaining approximately 490 images are used for training (Huang et al., 22 Sep 2025). Evaluation is based on photometric quality on held-out views.

The reported metrics are PSNR, SSIM, and LPIPS. The paper defines PSNR as

$$\mathrm{PSNR} \;=\; 10\,\log_{10}\!\bigl(\tfrac{L_{\max}^2}{\mathrm{MSE}}\bigr), \qquad \mathrm{MSE} = \tfrac{1}{HW}\sum_{x,y}\bigl(I_{xy} - \hat I_{xy}\bigr)^2.$$

SSIM is given as

$$\mathrm{SSIM}(I,\hat I)= \frac{(2\mu_I\mu_{\hat I}+C_1)\,(2\sigma_{I\hat I}+C_2)} {(\mu_I^2+\mu_{\hat I}^2+C_1)\,(\sigma_I^2+\sigma_{\hat I}^2+C_2)}.$$

LPIPS is used as described in Zhang et al. (Huang et al., 22 Sep 2025).

This evaluation design reflects a common distinction in underwater reconstruction research: held-out photometric consistency can be measured directly from images, whereas geometric fidelity remains only indirectly observable when no ground-truth depth or mesh is available. A plausible implication is that improvements on BlueCoral3D should be interpreted primarily as advances in novel-view synthesis quality and relative reconstruction behavior rather than definitive evidence of metrically accurate 3D recovery.

5. Baseline performance and comparative results

Table 2 of the paper reports scene-averaged quantitative performance for Scene 1 and Scene 2 across SeaThru-NeRF, 3DGS, UW-GS, WaterSplatting, WildGaussian, and R-Splatting (Huang et al., 22 Sep 2025).

Method	Scene 1: PSNR / SSIM / LPIPS	Scene 2: PSNR / SSIM / LPIPS
SeaThru-NeRF	15.92 / 0.437 / 0.604	20.66 / 0.619 / 0.541
3DGS	21.17 / 0.772 / 0.285	24.45 / 0.848 / 0.271
UW-GS	21.33 / 0.752 / 0.290	24.96 / 0.851 / 0.265
WaterSplatting	20.80 / 0.721 / 0.348	24.45 / 0.832 / 0.299
WildGaussian	21.76 / 0.748 / 0.387	24.98 / 0.849 / 0.304
R-Splatting	23.04 / 0.780 / 0.301	26.31 / 0.860 / 0.253

These results show that R-Splatting outperforms both classical 3DGS and specialized underwater NeRF baselines by approximately 1–2 dB in PSNR and achieves modest gains in SSIM and LPIPS (Huang et al., 22 Sep 2025). The dataset therefore functions not only as an image collection but as a benchmark exposing the interaction between underwater restoration, illumination variability, and 3D Gaussian Splatting-based reconstruction.

6. Relation to coral-reef photogrammetry and semantic reconstruction

BlueCoral3D is centered on shallow reef scenes, and the broader coral-reef imaging literature provides methodological context for what it does and does not include. “Combining Photogrammetric Computer Vision and Semantic Segmentation for Fine-grained Understanding of Coral Reef Growth under Climate Change” describes a distinct pipeline in which underwater video frames are processed through radiometric color correction, Root-SIFT feature extraction, AdaLAM matching, GCP-constrained bundle adjustment, dense reconstruction, Poisson surface reconstruction, orthophoto and DSM generation at 1 mm resolution, DeepLabv3+-based semantic segmentation with ASPP and ShapeConv, and reprojection of 2D labels into 3D space (Zhong et al., 2022).

That pipeline reports detailed 3D mesh models, digital surface models and orthophotos, uses five ground control points with known geodetic coordinates, and achieves orthophoto and DSM resolution of 1 mm/pixel with RMSE < 5 mm in all three axes (Zhong et al., 2022). It also reports semantic metrics such as mPA and mIoU, along with height-change and rugosity analyses.

The contrast is informative. BlueCoral3D, as described in the underwater 3DGS paper, does not include depth maps, meshes, dense point clouds, semantic labels, or auxiliary physical metadata (Huang et al., 22 Sep 2025). By comparison, the coral-reef photogrammetry workflow in (Zhong et al., 2022) exemplifies a richer ecological surveying pipeline with georeferenced geometry and semantic analysis. This suggests a useful conceptual distinction between a rendering-oriented benchmark and a monitoring-oriented reconstruction framework.

7. Limitations, ambiguities, and future extensions

The authors state “We release BlueCoral3D,” but no download link or DOI appears in the paper, and licensing terms are not specified (Huang et al., 22 Sep 2025). Interested users are encouraged to contact the authors or check the project webpage once published.

The dataset deliberately targets dynamic, caustic-lit shallow reefs. Other underwater conditions, including turbid deep water, shipwreck interiors, and synthetic tank benchmarks, remain unaddressed (Huang et al., 22 Sep 2025). The absence of ground-truth 3D geometry limits evaluation of absolute geometric accuracy. Future extensions might integrate sonar or structured-light scans, or control-tank captures with known turbidity parameters. The lack of semantic labels and physical metadata also precludes research into joint semantic-photometric reconstruction or physics-informed rendering.

A further point of clarification concerns nomenclature. The provided material pairs the BlueCoral3D dataset description from (Huang et al., 22 Sep 2025) with a separate coral-reef reconstruction and semantic analysis pipeline from (Zhong et al., 2022). This suggests a potential naming ambiguity rather than a single uniformly defined benchmark. For technical usage, the most precise interpretation is to treat BlueCoral3D primarily as the two-scene underwater dataset introduced for evaluating R-Splatting, while regarding the coral photogrammetry work as a related methodological reference rather than a definitive specification of the same dataset.

In summary, BlueCoral3D is a compact but technically focused underwater benchmark for multi-view image-based reconstruction in shallow coral reef environments under naturally varying illumination. Its strengths are calibrated 360° image sweeps, high-resolution imagery, and a clear held-out-view protocol; its principal limitations are narrow environmental coverage, missing 3D ground truth, and sparse metadata (Huang et al., 22 Sep 2025). Within those bounds, it provides a testbed for studying the coupling of underwater image restoration, illumination variation, and 3D Gaussian Splatting-based reconstruction.