EREBUS: End-to-end Robust Event Based Underwater Simulation

Hitesh Kyatham, Arjun Suresh, Aadi Palnitkar, Yiannis AloimonosarXiv 2025 · DOI · arXiv

Introduction

Underwater robotics faces a "perfect storm" of visual challenges, including low light, high dynamic range, and turbidity caused by suspended particles. Traditional frame-based cameras often fail in these conditions due to motion blur and limited temporal resolution. Event cameras, or Dynamic Vision Sensors (DVS), offer a promising alternative by recording only asynchronous brightness changes at each pixel. However, the adoption of event cameras in marine environments is hindered by a severe lack of labeled real-world datasets.

EREBUS (End-to-end Robust Event Based Underwater Simulation) addresses this gap by providing a novel simulation pipeline designed to generate realistic event-based data for challenging underwater scenes. By leveraging physics-based rendering, the framework enables the training of robust computer vision models on synthetic data, facilitating a "sim-to-real" transition for underwater perception.

The EREBUS Pipeline

The EREBUS pipeline consists of three integrated stages designed to mimic the complexities of the ocean floor:

  1. Underwater Scene Simulation: Using Blender, researchers create 3D environments featuring randomized seabed rock formations and thousands of suspended particles to simulate "marine snow" or silt. Lighting is provided by a virtual spotlight attached to an Autonomous Underwater Vehicle (AUV).
  2. Event Stream Generation: Rendered RGB frames and segmentation masks are processed through an event camera simulator. This converts intensity changes between frames into a time-ordered stream of events $(x, y, \text{polarity}, \text{timestamp})$.
  3. Perception Training: The resulting event data is used to train a YOLOv8-n (nano) model. This lightweight architecture is specifically chosen for its suitability on resource-constrained underwater edge devices.
The EREBUS pipeline stages: from dark RGB inputs to DVS event generation and final model prediction.

Key Insights and Experimental Results

A primary technical contribution of EREBUS is its Adaptive Training approach. By systematically varying the size and density of simulated particles, the researchers can test and improve model robustness against visual noise.

High Efficiency and Robustness

The study revealed that event data is remarkably information-dense. The YOLOv8-n model achieved a Mean Average Precision (mAP) of 0.83 using a training set of only 10 annotated images. This suggests that synthetic event data may be significantly more "efficient" for training than traditional synthetic RGB data in low-visibility environments.

Quantifying Noise Impact

The simulation allowed for precise control over "particle noise." Researchers discovered that when simulated particles exceeded twice the baseline size, DVS performance degraded significantly due to background clutter. This mapping of noise impact provides a roadmap for developing future noise-filtering algorithms specifically for underwater event streams.

Discussion and Future Work

While EREBUS currently focuses on rock detection, it serves as a foundational tool for underwater robotics. Future iterations will expand to more complex biological scenes, such as coral reefs and moving marine life.

The authors advocate for a hybrid sensing approach, where event sensors are integrated to handle specific "edge cases" like rapid motion and extreme low light where frame-based systems typically fail. The team intends to release the EREBUS pipeline as an open-source tool and provide a foundational public benchmark dataset to foster broader research in the underwater event-vision community.

Summary updated

BibTeX
@misc{kyatham2025erebus,
  author = {Hitesh Kyatham and Arjun Suresh and Aadi Palnitkar and Yiannis Aloimonos},
  title = {{EREBUS: End-to-end Robust Event Based Underwater Simulation}},
  year = {2025},
  eprint = {2511.01381},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.01381},
}
Figure