RealSense-Based Data Collection for Underwater Experiments: An Open Resource for the Research Community

Underwater sensing is one of those domains where the gap between controlled laboratory conditions and real-world performance can be enormous. Light behaves differently underwater. Turbidity scatters signals. Noise creeps into every measurement. For researchers and engineers working on underwater perception, having access to well-documented, controlled datasets is invaluable — and yet such resources remain surprisingly scarce.

 

That's why we're excited to share our open repository for RealSense-based underwater data collection, built around controlled water tank experiments using Intel RealSense depth cameras.

 

Anyone who has tried to use an off-the-shelf depth sensor underwater knows the frustration. Cameras designed for air don't simply work when submerged. Water attenuates light, particularly at certain wavelengths. Suspended particles scatter infrared signals that structured-light and stereo sensors rely on. The result is degraded depth maps, noisy point clouds, and unreliable distance measurements.

 

Understanding exactly how and why this degradation happens — and developing methods to compensate for it — requires systematic data collection under controlled conditions. That's precisely what this project delivers.

Our repository provides everything needed to replicate and extend our experimental setup. At its core is a data acquisition pipeline built around the Intel RealSense camera family (D435 and D455 models), deployed in a water tank where experimental conditions can be carefully managed.

 

The repository includes the data acquisition scripts for capturing synchronised colour and depth streams from the RealSense hardware, sample datasets collected across various water tank conditions, and supporting tools for data handling and visualisation. The entire pipeline is built on accessible, well-supported tools — Python, the Intel RealSense SDK, OpenCV, and NumPy — making it straightforward for other researchers to adopt and adapt.

Real ocean or lake environments are inherently unpredictable. Currents shift, visibility changes by the hour, and biological activity introduces variables that are nearly impossible to isolate. A water tank strips away that complexity, allowing researchers to study specific phenomena — the effect of turbidity on depth accuracy, for instance, or how light attenuation varies with distance — in a repeatable, measurable way.

 

This controlled approach doesn't replace field testing, but it provides the foundation for it. By first characterising sensor behaviour in known conditions, researchers can build models and correction algorithms that are then validated in open water.

This work connects to our broader research programme in underwater perception and marine monitoring. Our published work on enhancing underwater situational awareness through deep learning has demonstrated that combining RealSense data with modern machine learning can significantly improve depth perception and distance measurement in challenging underwater environments. We've also explored how multimodal sensing approaches can support environmental protection and the smart energy transition through more effective marine monitoring.

 

By making our data collection tools and datasets openly available, we hope to lower the barrier for other research groups working on similar problems — whether in underwater robotics, marine inspection, aquaculture monitoring, or subsea infrastructure assessment.

We encourage the community to use, extend, and build upon this work. If it proves useful in your research, we'd appreciate a citation — and we'd love to hear what you're working on.

 

If you find this work useful in your research, please consider citing the following publications:

 

@inproceedings{tolie2024enhancing,
title = {Enhancing Underwater Situational Awareness: RealSense Camera Integration with Deep Learning for Improved Depth Perception and Distance Measurement},
author = {Tolie, Hamidreza Farhadi and Ren, Jinchang and Hasan, Md Junayed and Kannan, Somasundar},
booktitle = {Artificial Intelligence for Security and Defence Applications II},
volume = {13206},
pages = {34--42},
year = {2024},
publisher = {SPIE}
}

 

@article{farhadi2024effective,
title = {Effective Marine Monitoring with Multimodal Sensing and Improved Underwater Robotic Perception towards Environmental Protection and Smart Energy Transition},
author = {Farhadi Tolie, Hamidreza and Ren, Jinchang and Hasan, Md Junayed and Ma, Ping and Kannan, Somasundar and Li, Yinhe},
journal = {Journal of Geodesy \& Geoinformation Science},
volume = {7},
number = {4},
year = {2024}
}