The Marine Debris Forward-Looking Sonar Datasets
Matias Valdenegro-Toro, Deepan Chakravarthi Padmanabhan, Deepak Singh, Bilal Wehbe, Yvan Petillot · 2025
Read The Marine Debris Forward-Looking Sonar Datasets by Matias Valdenegro-Toro, Deepan Chakravarthi Padmanabhan, Deepak Singh, Bilal Wehbe, Yvan Petillot (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-22880.txt.
Introduction
Matias Valdenegro-Toro and co-authors release data sets of sonar images of underwater trash, in this 2025 paper. Sonar is the main sense for robots working below the surface, where cameras fail. AI tools for sonar lag behind, in part because public training data are scarce. Their sets aim to close that gap for the task of finding marine debris.
Images were gathered in three settings of rising realism: a water tank, a turntable setup, and a flooded quarry. The sets support several tasks, from naming objects to outlining them pixel by pixel. The paper describes the data, gives first test results, and links a public archive. Robots that can spot trash are a step toward cleaning harbors and coasts.
Catalog record
| CREATOR | Matias Valdenegro-Toro, Deepan Chakravarthi Padmanabhan, Deepak Singh, Bilal Wehbe, Yvan Petillot |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-22880
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The Marine Debris Forward-Looking Sonar Datasets Matias Valdenegro-Toro, Deepan Chakravarthi Padmanabhan, Deepak Singh, Bilal Wehbe, Yvan Petillot 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 . I Introduction Machine learning has provoked many advances in the interpretation and processing of sonar images, but these techniques are driven by the availability of large datasets, which are not always publicly available. The machine learning and computer vision communities have many benchmark datasets that can be used for rapid prototyping and testing of new algorithms and methods, like ImageNet [ 1 ] and OpenImages [ 2 ] . Research in underwater perception and sonar image processing has similar requirements [ 3 ] , but datasets are usually not publicly available due to military use or other issues. This hinders improvements in the field due to lack of standard benchmarks, and increases the barrier to newcomers into the field. Transfer learning has been a standard approach to improve performance in computer vision systems [ 4 ] for a long time, but these improvements are problematic in the acoustic imaging domain, mostly due to the lack of pre-trained models trained specifically for sonar images, and more importantly, large datasets for unsupervised and self-supervised learning. Fuchs et al. [ 5 ] uses off the shelf models pre-trained on color images (ImageNet specifically), showing that these also improve performance in sonar image classification, but these improvements are only possible due to using pre-training on ImageNet (with > 1M images). Most research in sonar image classification is about detecting objects of military interests, such as marine mines or wreckage, but there are many possible civilian applications. In particular we focus on human made debris sitting on the seafloor, such as plastic, metal cans, bottles, marine nets, etc. In this paper we propose a new publicly available forward-looking sonar dataset, consisting of marine debris objects in multiple settings, and multiple tasks. The Marine Debris FLS dataset covers image classification, object detection, semantic segmentation, patch matching, and unsupervised feature learning, among other possible use cases and applications. The dataset was captured in a laboratory (Watertank and Turntable) and real-world (Flooded Quarry) settings using a ARIS Explorer 3000 Forward-Looking Sonar, and can be used as basis for future research in understanding sonar images and development of artificial intelligence models. This paper describes the dataset in detail, including design decisions, data capture, labeling, and some initial benchmarks. (a) Scene setup. (b) FLS image. Fig. 1 : Example scene and sonar image in the watertank dataset. Contributions of this paper are: • The marine debris datasets in three different settings and up to four different computer vision tasks. • One of the first publicly available sonar image datasets. • Initial benchmark results showing the usefulness of our datasets. Parts of the FLS datasets have been partially described in previous publications, like [ 6 ] for the FLS segmentation dataset, and [ 7 ] for the FLS quarry dataset, and [ 8 ] for the Turntable dataset. Figures 2 , 3 , and 4 visualizes the three datasets and provides basic visual information. Watertank Turntable Quarry Labels Object Bounding Boxes, Semantic Segmentation, Class Labels Class Labels, Material Labels None Classes Bottle, Can, Chain, Drink Carton, Hook, Propeller, Shampoo Bottle, Standing Bottle, Tire, Valve, Background (11) Bottle, Can, Drink Carton, Box, Bidon, Pipe, Platform, Propeller, Sachet, Tire, Valve, and Wrench (12) None Tasks Object Detection, Semantic Segmentation, Image Classification, Patch Matching Object Classification, Material Classification Feature Learning Number of Samples 2627 2150 7209 Examples per Task OC : 2627, OD : 1868, SS : 1868, PM : 47280 OC : 2150, MC : 2150 UL : Variable Variability Low Medium High Sample Images Image Format Full Size and Patches Patches Full Size and Patches TABLE I : Overview of Datasets and their related tasks: Object Classification (OC), Object Detection (OD), Patch Matching (PM), Semantic Segmentation (SS). In this paper we publicly release and unify all datasets into a single large dataset with multiple settings and multiple tasks, as shown in Table I . For the Watertank dataset we have object detection and patch matching labels that are newly released, while the Turntable dataset has material recognition labels, we make full descriptions of the Turntable dataset including platform and object crops, and the Quarry dataset we release the full set of sonar frames (7209 in total) and we describe how to extract patches that are useful for unsupervised and self-supervised learning. II Sonar Data Capture This section describes how data was captured, the three scenarios, and which object classes were used in each dataset. Sensor Selection . With the objective of detecting marine debris in the seafloor, selecting an appropriate sonar sensor is crucial due to object size. Most marine debris objects are only centimetres in size (i.e. bottles, cans, drink packaging, etc), while standard sensors like sidescan sonar provide per-pixel resolutions around 10 cm per pixel. Forward-Looking Sonar (FLS) sensors usually have higher per-pixel resolutions, in particular we selected the ARIS Explorer 3000 [ 9 ] as it provides resolutions up to 2.3mm per pixel (at the 3.0 MHz setting) and frame rates up to 15 FPS. Marine debris objects can clearly be seen in the output from this sensor as shown in Figure 1 . Watertank Scenario . As starting data, we captured a dataset the Water Tank of the Ocean Systems Laboratory, Heriot-Watt University, Edinburgh, Scotland. The Tank measures approximately (W,H,D)=3\times 2\times 4 [m], and on it we submerged the Nessie AUV with a sonar sensor attached to the underside, as shown in Figure 1 . Objects were placed in the water tank bottom in a pseudo-random way, without overlapping each other. Positioning was performed by dropping them into the water, and adjusting position using a pole manually. The Nessie AUV was tele-operated to capture different views of the scene, and the water tank size limited the range of possible movements. This constraints the field of view and object poses. The vehicle performed a quarter circle sweep from side to side in order to capture all objects in the tank floor, in multiple starting positions inside the water tank. We selected a set of household and typical objects found in marine environments, motivated that most debris objects are composed of discarded man-made objects. We also included a set of set of marine objects as distractors, which are also expected to be found in marine environments but they are not necessarily marine debris. We describe these sets below: Household objects . Bottles of different materials (glass, metal, plastic), tin cans, drink cartons (milk and juice), plastic shampoo bottles, a metal box, a plastic bidon, a drink sachet, and a metal wrench. Distractors . Chain, hook, propeller, rubber tire (two sizes),