(Left) Raw video by Piyapong Suwannakul and (Right) Processed video using VI-Lab tools. See video HERE
Our oceans have been explored for hundreds of years and these activities are becoming increasingly important because of the need to manage and conserve mineral and biological resources effectively, as well as to better understand planetary-scale processes including tectonics and marine hazards. Exploration and analysis are however always limited by the number of diving experts, technologies, and in particular, costs. Advanced imaging methods now support a new paradigm of remote discovery where onshore experts with specific knowledge, such as geologists, archaeologists and biologists, are able to remotely model and explore underwater scenes.
Underwater environment represents the combination of several challenges. Water is a dynamic medium and suspended particles move. Light scatter causes blur and halo effects, whilst light absorption leads to colour distortion and reduced contrast. The model of underwater imagery should thus comprise temporally- and spatially-variant distortion, uneven intensity bias, multiplicative noise, and additive noise. This project aims to exploit underwater image priors to perform 3D mapping process can be done directly from the raw underwater sequences.
[
Our enhanced Gaussian Splatting framework improves both visual quality and geometric accuracy in underwater rendering. We employ physics-guided decoupled RGB learning for accurate colour restoration, a frame interpolation strategy with adaptive weighting to address sparse views, and a new loss function that reduces noise while preserving edges, crucial for deep-sea content.
We present a semantic-guided 3D Gaussian Splatting framework for deep-sea scene reconstruction, where each Gaussian embeds CLIP-derived features to enforce semantic and structural consistency. A dedicated semantic loss and stage-wise training strategy further enhance stability and reconstruction fidelity.
[Project webpage] [PDF
Our Gaussian Splatting-based method introduces a color appearance model for distance-dependent color variation, employs a new physics-based density control strategy to enhance clarity for distant objects, and uses a binary motion mask to handle dynamic content. The method is optimised with a well-designed loss function supporting scattering media and strengthened by pseudo-depth maps.
The framework introduces a novel method for generating paired datasets from raw underwater videos, producing snowy and snow-free pairs that enable supervised training for video enhancement.