Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Marine Debris Forward-Looking Sonar Datasets

Matias Valdenegro-Toro, Deepan Chakravarthi Padmanabhan, Deepak Singh, Bilal Wehbe, Yvan Petillot

TL;DR

This paper addresses the lack of public sonar benchmarks by introducing the Marine Debris Forward-Looking Sonar datasets, collected across three settings (Watertank, Turntable, Quarry) to provide diverse acoustic views. It unifies multiple computer vision tasks—Object Classification, Object Detection, Semantic Segmentation, Patch Matching, and Unsupervised Learning—along with detailed labeling, preprocessing, and data organization. The authors present initial benchmarks that reveal task difficulty and the current limitations of transfer learning and self-supervised methods in sonar, while showcasing the Quarry dataset as a large unsupervised resource. Public availability aims to standardize evaluation, accelerate progress in underwater perception, and spur sonar-specific learning approaches with practical impact for robotics and environmental monitoring.

Abstract

Sonar sensing is fundamental for underwater robotics, but limited by capabilities of AI systems, which need large training datasets. Public data in sonar modalities is lacking. This paper presents the Marine Debris Forward-Looking Sonar datasets, with three different settings (watertank, turntable, flooded quarry) increasing dataset diversity and multiple computer vision tasks: object classification, object detection, semantic segmentation, patch matching, and unsupervised learning. We provide full dataset description, basic analysis and initial results for some tasks. We expect the research community will benefit from this dataset, which is publicly available at https://doi.org/10.5281/zenodo.15101686

The Marine Debris Forward-Looking Sonar Datasets

TL;DR

This paper addresses the lack of public sonar benchmarks by introducing the Marine Debris Forward-Looking Sonar datasets, collected across three settings (Watertank, Turntable, Quarry) to provide diverse acoustic views. It unifies multiple computer vision tasks—Object Classification, Object Detection, Semantic Segmentation, Patch Matching, and Unsupervised Learning—along with detailed labeling, preprocessing, and data organization. The authors present initial benchmarks that reveal task difficulty and the current limitations of transfer learning and self-supervised methods in sonar, while showcasing the Quarry dataset as a large unsupervised resource. Public availability aims to standardize evaluation, accelerate progress in underwater perception, and spur sonar-specific learning approaches with practical impact for robotics and environmental monitoring.

Abstract

Sonar sensing is fundamental for underwater robotics, but limited by capabilities of AI systems, which need large training datasets. Public data in sonar modalities is lacking. This paper presents the Marine Debris Forward-Looking Sonar datasets, with three different settings (watertank, turntable, flooded quarry) increasing dataset diversity and multiple computer vision tasks: object classification, object detection, semantic segmentation, patch matching, and unsupervised learning. We provide full dataset description, basic analysis and initial results for some tasks. We expect the research community will benefit from this dataset, which is publicly available at https://doi.org/10.5281/zenodo.15101686

Paper Structure

This paper contains 13 sections, 12 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Sonar Data Capture
  3. Marine Debris FLS Datasets
  4. Tasks
  5. Labeling
  6. Preprocessing
  7. File Structure and Splits
  8. Licensing
  9. Initial Results and Benchmarks
  10. Conclusions and Future Work
  11. Broader Impact Statement
  12. Additional Dataset Information and Statistics
  13. Additional Failed FLS Images from Quarry Scenario

Figures (12)

  • Figure 1: Example scene and sonar image in the watertank dataset.
  • Figure 2: Visual description of the Watertank scenario, including the AUV setup, objects in the watertank, captures FLS images, and a Multidimensional Scaling visualization of patches, showing the different classes and their pixel-space neighbors.
  • Figure 3: Visual description of the Turntable Scenario, including the Turntable setup and object placement, sample FLS images as the turntable rotates, and MDS visualization of patches, showing the increased diversity of this dataset.
  • Figure 4: Visual description of the Quarry scenario, including fullsize FLS images, premade patch dataset statistics, MDS visualization, and some sample patches.
  • Figure 5: Sample color and FLS images from the watertank dataset, showcasing the real-world object that were used to build this dataset.
  • ...and 7 more figures