💻 computer science

WaterSplat-SLAM: Photorealistic Monocular SLAM in Underwater Environment

WaterSplat-SLAM is a novel monocular underwater SLAM system that achieves robust pose estimation and photorealistic dense mapping by integrating semantic medium filtering for improved tracking and depth estimation, alongside a semantic-guided rendering and adaptive map management strategy using online medium-aware Gaussian maps.

Original authors: Kangxu Wang, Shaofeng Zou, Chenxing Jiang, Yixiang Dai, Siang Chen, Shaojie Shen, Guijin Wang

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kangxu Wang, Shaofeng Zou, Chenxing Jiang, Yixiang Dai, Siang Chen, Shaojie Shen, Guijin Wang

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 take a photo underwater. You know the drill: the water acts like a foggy, blue-tinted curtain. It scatters light, makes things look blurry, and washes out colors. Now, imagine trying to build a 3D map of the ocean floor using just a single camera, like a diver's GoPro. That is the incredibly difficult puzzle this paper solves.

The authors introduce WaterSplat-SLAM, a new system that lets robots "see" clearly underwater and build a photorealistic 3D map in real-time. Here is how it works, broken down with some everyday analogies.

The Problem: The "Foggy Window" Effect

Traditional underwater mapping is like trying to draw a detailed map of a room while looking through a dirty, foggy window.

Old methods either gave up and made a blurry, low-resolution sketch (using sonar) or tried to use standard 3D mapping tools that got confused by the water, thinking the "fog" was actually part of the objects (like a rock or a fish).
The result: The maps were either too blocky to be useful or looked nothing like the real world.

The Solution: WaterSplat-SLAM

Think of WaterSplat-SLAM as a smart, magical painter that knows exactly how to separate the "fog" from the "objects." It uses a technique called 3D Gaussian Splatting, which is like building a scene out of millions of tiny, fluffy, colored clouds (Gaussians) instead of rigid bricks.

Here are the three "superpowers" that make it work:

1. The "Smart Goggles" (Semantic Filtering)

When a robot looks underwater, it sees a mix of rocks, fish, and the water itself. Standard cameras get confused and try to map the water particles as if they were solid rocks.

The Analogy: Imagine wearing a pair of smart glasses that can instantly tell the difference between a person and the air around them.
How it works: The system uses an AI "goggle" to identify the water in the image and says, "Ignore this part; that's just the medium." It filters out the water's visual noise so the robot can focus on tracking its position relative to the actual objects (rocks, pipes, ruins). This prevents the robot from getting lost in the "fog."

2. The "Two-Part Paintbrush" (Medium-Aware Rendering)

Once the robot knows where it is, it needs to paint the 3D map.

The Analogy: Imagine painting a scene where you have to paint the objects (like a coral reef) and the atmosphere (the blue, murky water) separately. If you mix them, the coral looks weird.
How it works: Instead of just painting "things," this system has a special brush that understands water physics. It predicts three things about the water: how much it dims the light (attenuation), how much it scatters light back (backscatter), and what color the water is.
The Magic: It renders the scene by calculating: "Here is the object, and here is the blue fog sitting in front of it." This allows it to create a map that looks incredibly realistic, preserving the eerie blue glow of the deep ocean while keeping the details of the objects sharp.

3. The "Tidy-Up Crew" (Adaptive Management)

As the robot moves and loops back around (like walking in a circle), it might draw the same object twice or create overlapping "clouds" of data, making the map huge and slow.

The Analogy: Imagine you are drawing a map on a whiteboard. Every time you walk past a spot, you add a new layer of marker. Eventually, the board is a mess. A "Tidy-Up Crew" comes in, sees the overlapping layers, and smudges them into one clean, perfect line.
How it works: When the robot realizes it has returned to a spot it has seen before (Loop Closure), the system merges the overlapping 3D "clouds" into single, efficient ones. This keeps the map small, fast, and accurate, even for long missions.

Why Does This Matter?

Before this, underwater robots were like blind explorers or artists with a shaky hand. They couldn't build high-quality, photo-realistic maps of the ocean floor.

WaterSplat-SLAM changes the game by:

Seeing through the fog: It ignores the water's distortion to track movement accurately.
Painting the truth: It creates maps that look like real photos, not just blocky 3D models.
Working in real-time: It's fast enough for a robot to use while it's actually swimming, not just after the fact.

In short: It's like giving an underwater robot a pair of high-definition, fog-piercing glasses and a magic paintbrush that knows exactly how to paint the ocean, allowing it to explore shipwrecks, inspect pipelines, and study marine life with crystal-clear vision.

1. Problem Statement

Underwater monocular Simultaneous Localization and Mapping (SLAM) is critical for applications like marine archaeology and autonomous underwater vehicle (AUV) navigation. However, existing methods face significant challenges:

Optical Degradation: Underwater scenes suffer from light attenuation and scattering, causing color shifts, blur, and low contrast.
Limitations of Existing SLAM:
- Feature-based methods fail in low-texture, turbid, or poorly lit environments.
- Dense mapping methods often rely on multi-sensor fusion (sonar, laser) which lacks photometric detail or produces sparse/occupancy maps unsuitable for photorealistic rendering.
- 3D Gaussian Splatting (3DGS) SLAM: While successful on land, standard 3DGS treats the scene as geometric primitives. In underwater environments, it fails to model the water medium itself, leading to artifacts where the scattering medium is incorrectly represented by geometric Gaussians, degrading tracking stability and rendering fidelity.

2. Methodology: WaterSplat-SLAM

The authors propose WaterSplat-SLAM, a monocular SLAM system that integrates semantic understanding with 3D Gaussian Splatting to model both underwater objects and the surrounding water medium simultaneously.

A. System Overview

The system takes an RGB sequence as input and outputs a photorealistic, medium-aware Gaussian map. The pipeline consists of three main stages:

Semantic Medium Filtering: A segmentation model (CLIPSeg) generates binary masks to identify water regions.
Camera Tracking & Initialization: Uses a two-view 3D reconstruction module (based on MASt3R) to estimate camera poses and pointmaps. Crucially, it excludes water pixels from the tracking optimization to prevent water-induced noise from corrupting pose estimation.
Online Medium-Aware Gaussian Mapping: Initializes and optimizes Gaussian primitives. It employs a Medium Network to predict water properties and a Semantic-Guided Rendering strategy to separate object rendering from medium rendering.

B. Key Technical Components

1. Semantic-Guided Monocular Tracking

Water Masking: The system uses CLIPSeg to generate water masks ( $M$ ).
Confidence Suppression: During the matching of pointmaps between frames, the confidence scores of water pixels are zeroed out ( $Q' = Q \odot (1-M)$ ). This ensures the camera pose estimation relies solely on solid objects, avoiding drift caused by the shifting water medium.

2. Online Medium-Aware Gaussian Map

Dual Rendering Model: Unlike standard 3DGS, WaterSplat-SLAM explicitly separates the rendering of objects and the water medium.
- Object Rendering ( $\hat{C}_{obj}$ ): Standard Gaussian splatting for solid objects.
- Medium Rendering ( $\hat{C}_{med}$ ): A Medium Network (MLP) predicts three parameters based on ray direction: attenuation density ( $\sigma_{attn}$ ), backscatter density ( $\sigma_{bs}$ ), and medium color ( $c_{med}$ ).
- Final Image: $\hat{C}_{render} = \hat{C}_{obj} + \hat{C}_{med}$ .
Semantic-Guided Loss: The optimization loss is split based on the water mask:
- For water pixels, the loss minimizes the difference between the rendered medium and the ground truth, preventing the system from trying to fit geometric Gaussians to the water.
- For object pixels, standard photometric and geometric losses are applied.
- An isotropic scale loss is applied only to visible primitives to prevent geometric distortion.

3. Adaptive Map Management

Loop Closure & Merging: Upon detecting loop closure, the system performs global bundle adjustment. To reduce redundancy and storage, Gaussian primitives bound to overlapping keyframes are merged.
Voxel Downsampling: Primitives within the same 3D voxel are merged into a single representative primitive (averaging color, scale, orientation, position, and opacity). This significantly reduces map size without sacrificing rendering quality.

3. Key Contributions

Novel Monocular Underwater SLAM: A system capable of robust pose estimation and high-fidelity dense mapping using only RGB input.
Medium-Aware Gaussian Representation: An online Gaussian map that explicitly models the water medium via a learned network, avoiding the misrepresentation of volumetric water as geometric primitives.
Semantic-Guided Strategy: The integration of semantic masks to filter water during tracking and to guide the rendering loss, ensuring the medium and objects are modeled correctly.
Adaptive Management: A loop-closure-based merging strategy that reduces Gaussian primitive count by ~24% while maintaining global consistency and rendering quality.

4. Experimental Results

The system was evaluated on the SeaThru-NeRF dataset (four real-world underwater sequences) and a custom WaterSplat-SLAM dataset (collected in a pool with varying turbidity).

Rendering Quality:
- Outperformed state-of-the-art NeRF-based (GO-SLAM, GLORIE-SLAM) and 3DGS-based (MonoGS, S3PO-GS) methods.
- Achieved an average PSNR of 24.14 on SeaThru-NeRF and 30.19 on the custom dataset, significantly higher than competitors.
- Demonstrated superior detail in turbid water, correctly rendering both foreground objects and the background medium.
Tracking Accuracy (ATE):
- Achieved the lowest Absolute Trajectory Error (ATE) across all sequences (Avg: 0.219m).
- Competitors like ORB-SLAM3 and MonoGS failed or drifted significantly in highly turbid sequences due to feature degradation or lack of depth priors.
Efficiency:
- Operates at 2–5 FPS on an RTX 4090, comparable to other real-time SLAM systems.
- The primitive merging strategy reduced the map size by 24.3% with negligible impact on reconstruction quality.

5. Significance

WaterSplat-SLAM addresses a critical gap in marine robotics by enabling photorealistic, dense 3D reconstruction in challenging underwater environments using only a single camera. By decoupling the water medium from geometric objects, it solves the "scattering problem" that plagues existing 3DGS methods. This advancement allows for more reliable autonomous navigation, precise environmental feedback for inspection tasks, and better data visualization for marine archaeology, all while maintaining the computational efficiency required for real-time deployment.