Rongchai Wang
May 31, 2025 16:36
Researchers have developed SeaSplat, an AI model that transforms underwater images into clear 3D scenes, aiding ecologists in studying coral reefs and their biodiversity.
In a groundbreaking development, researchers from the Woods Hole Oceanographic Institution (WHOI) and the Massachusetts Institute of Technology (MIT) have introduced an innovative AI model named SeaSplat. This model is designed to transform unclear underwater images into vivid, highly detailed 3D scenes, thereby providing ecologists with a more precise tool to study and understand underwater environments, including coral reefs.
Addressing Underwater Image Distortions
SeaSplat has been developed to counteract the common distortions found in underwater photography, primarily haze and discoloration. These distortions often obscure the true colors and details of underwater scenes. By reconstructing what these scenes would look like without water interference, SeaSplat can produce sharp images that accurately reflect the natural colors of underwater objects and marine life.
Innovative 3D Reconstruction Techniques
The model leverages a technique known as 3D Gaussian splatting, allowing it to generate precise 360-degree images from two-dimensional photographs. By using a vast collection of underwater images captured by both divers and robots, the researchers have refined the model to automatically correct for inherent water distortions, thus providing a true-to-life representation of underwater scenes.
Impact on Coral Reef Studies
SeaSplat holds significant potential for ecological studies, particularly in understanding coral reefs, which are crucial to marine biodiversity. According to the International Coral Reef Initiative, about 84% of the world’s reefs have suffered from harmful bleaching over the past two years. The model can assist in monitoring these ecosystems by quantifying biodiversity and detecting events such as coral bleaching or disease.
Utilizing Advanced Technology
The development of SeaSplat was supported by advanced technology, including NVIDIA Jetson Orin edge compute for capturing underwater imagery and NVIDIA L40 GPUs for the model’s initial training. The model can now be applied to images from standard underwater cameras, making it accessible for broad ecological applications.
Future Applications
Having already been used to enhance images from the U.S. Virgin Islands, the Red Sea, and Curaçao, researchers plan to make SeaSplat more generalizable and scalable for various underwater studies. This advancement promises to provide ecologists with enhanced tools for assessing and preserving marine biodiversity.
For additional details, you can explore the original article on the NVIDIA Developer Blog.
Image source: Shutterstock
