Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Marine Snow Removal Using Internally Generated Pseudo Ground Truth

Alexandra Malyugina, Guoxi Huang, Eduardo Ruiz, Benjamin Leslie, Nantheera Anantrasirichai

TL;DR

The paper tackles the challenge of marine snow degrading underwater vision and hindering SLAM by addressing the scarcity of paired training data. It introduces a dataset-generation pipeline that creates snow-affected and clean frame pairs from real underwater footage through frame blending, enabling supervised learning for enhancement. A video enhancement module based on the BVI-Mamba framework with state-space models and 2D SS2D is then used to improve frame quality while preserving feature integrity, boosting SLAM robustness. Empirical results on real Beam data show improved SLAM metrics and denser 3D reconstructions, demonstrating practical impact for underwater robotics and mapping tasks in snow-impacted environments.

Abstract

Underwater videos often suffer from degraded quality due to light absorption, scattering, and various noise sources. Among these, marine snow, which is suspended organic particles appearing as bright spots or noise, significantly impacts machine vision tasks, particularly those involving feature matching. Existing methods for removing marine snow are ineffective due to the lack of paired training data. To address this challenge, this paper proposes a novel enhancement framework that introduces a new approach for generating paired datasets from raw underwater videos. The resulting dataset consists of paired images of generated snowy and snow, free underwater videos, enabling supervised training for video enhancement. We describe the dataset creation process, highlight its key characteristics, and demonstrate its effectiveness in enhancing underwater image restoration in the absence of ground truth.

Marine Snow Removal Using Internally Generated Pseudo Ground Truth

TL;DR

The paper tackles the challenge of marine snow degrading underwater vision and hindering SLAM by addressing the scarcity of paired training data. It introduces a dataset-generation pipeline that creates snow-affected and clean frame pairs from real underwater footage through frame blending, enabling supervised learning for enhancement. A video enhancement module based on the BVI-Mamba framework with state-space models and 2D SS2D is then used to improve frame quality while preserving feature integrity, boosting SLAM robustness. Empirical results on real Beam data show improved SLAM metrics and denser 3D reconstructions, demonstrating practical impact for underwater robotics and mapping tasks in snow-impacted environments.

Abstract

Underwater videos often suffer from degraded quality due to light absorption, scattering, and various noise sources. Among these, marine snow, which is suspended organic particles appearing as bright spots or noise, significantly impacts machine vision tasks, particularly those involving feature matching. Existing methods for removing marine snow are ineffective due to the lack of paired training data. To address this challenge, this paper proposes a novel enhancement framework that introduces a new approach for generating paired datasets from raw underwater videos. The resulting dataset consists of paired images of generated snowy and snow, free underwater videos, enabling supervised training for video enhancement. We describe the dataset creation process, highlight its key characteristics, and demonstrate its effectiveness in enhancing underwater image restoration in the absence of ground truth.
Paper Structure (10 sections, 3 equations, 5 figures, 1 table)

This paper contains 10 sections, 3 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Dataset Generation Module
  5. Video Enhancement Module
  6. Experiments and Results
  7. Datasets and Training Setup
  8. Performance Metrics
  9. Results
  10. Conclusion

Figures (5)

  • Figure 1: Diagram of the proposed framework
  • Figure 2: First column: extracted snow mask; second column: ground truth image; last column: resulting snow-affected image with overlaid snow mask.
  • Figure 3: Subjective results on test set. First column: Raw input image with real marine snow; second column: result; last column: ground truth image.
  • Figure 4: Number of keypoints per frame (lower is better, indicating fewer snow particles). Smoothing was applied to the data for clarity.
  • Figure 7: 3D models (point cloud) reconstructed for sequence 2023$\_$000015$\_$000135 (top) and sequence 2022_005644_005744. Methods order (left to right): no enhancement, our method, BEM. The red arrow highlights key differences (missing parts or sparsity) of the enhanced video with respect to the non-enhanced baseline. The green box highlights improved reconstruction with respect to the baseline