BALTIC: A Benchmark and Cross-Domain Approach for 3D Reconstruction in Scattering and Lighting

Michele Grimaldi

Keywords: Marine Robotics, Structure-from-Motion, Un- derwater Vision, 3D Representations

Abstract:

Robust 3D reconstruction across varying environmen-
tal conditions remains a critical challenge for robotic perception,
especially when transitioning between air and water. We present
a controlled benchmark to systematically evaluate modern 3D
reconstruction methods under variations in medium and lighting.
Our setup features a custom water tank with a monocular camera
and HTC Vive tracker, providing accurate ground-truth poses. We
introduce BALTIC, a benchmark comprising 13 datasets spanning
two media (air and water) and three lighting conditions (ambient,
artificial, mixed), with variations in motion type, scanning pattern,
and initialization trajectory, producing a heterogeneous set of
sequences. We further explore cross-domain reconstruction by
augmenting underwater images with a minimal set of in-air
views captured under similar lighting. Structure-from-Motion
reconstruction using COLMAP is evaluated for both trajectory
and scene geometry, serving as input to Neural Radiance Fields
and 3D Gaussian Splatting methods. Reconstructions are assessed
against ground-truth trajectories and in-air references, and
rendered outputs are compared using perceptual and photometric
metrics. We also performed a color restoration evaluation to
assess radiometric reconstruction across domains. In controlled,
texture-consistent conditions, Gaussian Splatting with simple
preprocessing (e.g., white balance) can achieve results comparable
to specialized underwater methods, though its performance may
vary in more complex, heterogeneous real-world environments.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Paper Acceptance: No

Submission Number: 9