💻 computer science

BALTIC: A Benchmark and Cross-Domain Strategy for 3D Reconstruction Across Air and Underwater Domains Under Varying Illumination

This paper introduces BALTIC, a comprehensive benchmark and cross-domain strategy that evaluates 3D reconstruction methods across air and underwater environments under varying illumination, demonstrating that Gaussian Splatting with simple preprocessing can achieve performance comparable to specialized underwater methods in controlled settings.

Original authors: Michele Grimaldi, David Nakath, Oscar Pizarro, Jonatan Scharff Willners, Ignacio Carlucho, Yvan R. Petillot

Published 2026-04-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Michele Grimaldi, David Nakath, Oscar Pizarro, Jonatan Scharff Willners, Ignacio Carlucho, Yvan R. Petillot

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a perfect 3D model of a sunken shipwreck. You have a camera, and you swim around the ship taking pictures. But underwater, the water is murky, the light is weird, and colors look washed out. It's like trying to solve a jigsaw puzzle while wearing foggy glasses in a dimly lit room.

This paper introduces a new tool called BALTIC to help robots and computers learn how to build these 3D models better, even when the conditions are terrible.

Here is the breakdown of what they did, using some everyday analogies:

1. The Problem: The "Foggy Glasses" Effect

When you take photos in the air (on land), the light is clear, and colors are bright. Computers are great at turning these photos into 3D models. But underwater, it's a different story.

The Water acts like a filter: It eats away at colors (making everything look green or blue) and scatters light (making things look blurry).
The Result: When a computer tries to build a 3D model from underwater photos, it often gets confused. The model might fall apart, look like a ghost, or have huge holes in it.

2. The Solution: The "Swimming Pool Lab" (BALTIC)

To fix this, the researchers built a giant, controlled swimming pool (a water tank) in a lab. Think of this as a training gym for robots.

They created 13 different scenarios in this tank.
They changed the lighting (bright sun, artificial lamps, or a mix).
They changed the water (clear, or murky with "clouds" to simulate the ocean).
They moved the camera in different patterns (like a lawnmower going back and forth, or a bird flying freely).
The Secret Weapon: They used a high-tech tracking system (like the ones used in VR games) to know exactly where the camera was at every single moment. This gave them the "answer key" to see if the computer's 3D model was actually correct.

3. The Experiment: Testing the "Detectives"

They tested three different types of computer "detectives" (algorithms) to see which one could solve the 3D puzzle best:

The Old School Detective (COLMAP): Great on land, but gets easily confused underwater. It's like a detective who can only solve crimes in broad daylight.
The AI Artists (NeRF & Gaussian Splatting): These are newer, smarter methods that learn to "paint" the scene. They did better, but they still struggled when the water was too murky.
The Specialized Divers: They also tested methods specifically designed for underwater use.

4. The Big Discovery: The "Air-Sea Rescue"

This is the most exciting part. The researchers found a clever trick to help the underwater detectives.

Imagine you are trying to draw a picture of a rusty old anchor underwater, but you can barely see it.

The Trick: Before you dive, you take a few photos of the same anchor while it's sitting in the air (on dry land) under similar lighting.
The Magic: You feed those clear, dry photos into the computer along with the murky underwater photos.
The Result: The computer uses the clear "dry" photos as a map to understand the shape and color of the object. It then uses that knowledge to "fill in the blanks" of the blurry underwater photos.

It's like giving a detective a clear photo of the suspect's face before asking them to identify the suspect through a foggy window. Suddenly, the reconstruction becomes sharp, complete, and accurate.

5. The Verdict

Underwater is hard: Even the best computers struggle to build 3D models from underwater photos alone because the water hides the details.
Preparation pays off: If you take a few photos in the air first, you can dramatically improve the underwater 3D model.
Simple fixes help: Just adjusting the colors (white balance) and brightening the contrast in the photos before feeding them to the computer helps a lot.

Why Does This Matter?

This isn't just about making cool 3D models. This technology helps:

Marine Biologists: Map coral reefs without touching them.
Archaeologists: Explore shipwrecks without damaging them.
Engineers: Inspect underwater pipelines and oil rigs to find cracks or leaks.

In short: The paper says, "Don't try to solve the underwater puzzle alone. Bring a few clear photos from the surface as a cheat sheet, and you'll build a much better 3D map of the deep."

1. Problem Statement

Robust 3D reconstruction is critical for marine robotics, underwater archaeology, and infrastructure inspection. However, existing methods (Structure-from-Motion, NeRF, 3D Gaussian Splatting) are predominantly evaluated in controlled, in-air environments with stable lighting. They struggle significantly when transitioning to underwater or cross-domain (air-to-water) scenarios due to:

Visual Degradation: Light scattering, absorption, and turbidity reduce contrast and color consistency.
Geometric Distortions: Refraction at the air-water interface.
Illumination Variability: Shifts between natural (sunlight), artificial, and mixed lighting sources.
Lack of Benchmarks: Existing datasets often isolate single factors (e.g., motion or lighting) and lack systematic variation across media and illumination, making it difficult to evaluate reconstruction robustness or compare methods under realistic conditions.

2. Methodology

A. The BALTIC Benchmark

The authors introduced BALTIC, a controlled, open-source benchmark designed to systematically evaluate 3D reconstruction.

Setup: A custom water tank ( $2.2 \times 1 \times 0.8$ m) equipped with a monocular camera inside a dome port and an HTC Vive motion tracking system for accurate ground-truth pose estimation (mean error < 3mm translation, < 0.3° rotation).
Data Composition: 13 datasets (E1–E13) varying across:
- Medium: Air vs. Water (simulated Baltic Sea conditions using dye and Maalox for scattering).
- Lighting: Directional (Sunlight), Artificial, and Mixed.
- Trajectory: "Lawn-mower" (structured) vs. Free-form 3D scanning.
Calibration: A rigorous calibration scheme was employed to undistort images into a common pinhole representation, isolating radiometric challenges (scattering/absorption) from geometric refraction errors.

B. Evaluation Pipeline

The study evaluated a standard pipeline and advanced learning-based methods:

Sparse Reconstruction: Used COLMAP for Structure-from-Motion (SfM) to estimate camera poses and sparse point clouds.
Dense Reconstruction: Used COLMAP outputs as input for:
- Neural Radiance Fields (NeRF): Specifically Nerfacto and SeaThru-NeRF.
- 3D Gaussian Splatting (3DGS): Basic 3DGS, WaterSplat, and SeaSplat.
Cross-Domain Strategy: Investigated augmenting underwater sequences with a minimal set of in-air images captured under similar lighting to provide geometric and radiometric priors.

C. Metrics

Performance was assessed using:

Geometric: Absolute Trajectory Error (ATE), Relative Pose Error (RPE), mesh statistics (surface area, curvature), and point cloud alignment (Chamfer distance, ICP fitness).
Photometric/Perceptual: PSNR, SSIM, and LPIPS.
Color Accuracy: CIE Lab $\Delta E$ differences.

3. Key Contributions

BALTIC Benchmark: The first controlled dataset systematically varying medium (air/water), lighting (natural/artificial/mixed), and motion patterns, complete with accurate ground-truth trajectories.
Comprehensive Evaluation: A systematic comparison of SfM, NeRF, and 3DGS methods, highlighting their specific failure modes (e.g., fragmentation in scattering media) and robustness limits.
Cross-Domain Reconstruction Strategy: A novel finding that augmenting underwater sequences with a small number of in-air images significantly improves reconstruction quality, offering a practical, low-cost strategy for real-world deployments.

4. Key Results

A. SfM and Sparse Reconstruction

Air vs. Water: COLMAP performs reliably in air (dense, consistent). In water, performance degrades drastically due to feature loss, resulting in sparse point clouds, shorter feature tracks, and fragmented reconstructions (multiple submaps).
Preprocessing: Simple preprocessing (cropping to light cones, CLAHE contrast enhancement) improved underwater SfM registration, reducing fragmentation.
Trajectory Accuracy: Underwater sequences showed higher ATE and RPE compared to air. However, Cross-Domain (E14) reconstruction (E12 underwater + E6 in-air) reduced drift and improved alignment, proving that in-air views stabilize the trajectory.

B. Dense Reconstruction (NeRF & 3DGS)

Performance Ranking: 3D Gaussian Splatting (3DGS) with basic preprocessing (white balance) consistently achieved the highest PSNR and SSIM across most datasets, outperforming specialized underwater NeRF variants in many cases.
Lighting Impact: Artificial lighting (E12) generally yielded better results than sunlight-only underwater conditions (E8) due to controlled illumination mitigating scattering effects.
Cross-Domain Impact: The cross-domain dataset (E14) showed superior metrics compared to pure underwater data (E12), with higher PSNR/SSIM and lower LPIPS, confirming that in-air views stabilize geometry and improve perceptual fidelity.

C. Color Restoration

Specialized methods like WaterSplat and SeaSplat showed strong perceptual color correction (restoring contrast and chromatic info) but exhibited higher mean $\Delta E$ errors compared to SeaThru-NeRF, which provided smoother transitions.
Simple white balance preprocessing significantly improved PSNR but sometimes degraded perceptual metrics (LPIPS), highlighting the trade-off between pixel-wise fidelity and structural consistency.

5. Significance and Conclusion

Practical Strategy: The paper demonstrates that cross-domain augmentation is a highly effective, lightweight strategy. Acquiring a few in-air images before underwater deployment can serve as a "geometric anchor," significantly improving reconstruction completeness and trajectory stability without requiring complex physical models or retraining.
Method Robustness: Standard methods like 3D Gaussian Splatting, when paired with simple preprocessing, can rival specialized underwater methods in controlled environments, challenging the notion that complex physics-based models are always necessary.
Limitations & Future Work: The benchmark is currently limited to a controlled tank environment. Future work aims to scale to real-world in-situ environments, incorporate diverse water conditions, and integrate physics-based rendering models more tightly with geometric reconstruction.

In summary, BALTIC provides a reproducible framework to analyze the limitations of current 3D reconstruction in marine environments and validates that cross-domain data fusion is a critical, practical solution for overcoming the visual degradation inherent in underwater robotics.