Abstract
Marine environments present significant challenges for perception and autonomy due to dynamic surfaces, limited visibility, and complex interactions between aerial, surface, and submerged sensing modalities. This paper introduces the Aerial Marine Perception Dataset (AMP2026), a multi-platform marine robotics dataset collected across multiple field deployments designed to support research in two primary areas: multi-view tracking and marine environment mapping. The dataset includes synchronized data from aerial drones, boat-mounted cameras, and submerged robotic platforms, along with associated localization and telemetry information. The goal of this work is to provide a publicly available dataset enabling research in marine perception and multi-robot observation scenarios. This paper describes the data collection methodology, sensor configurations, dataset organization, and intended research tasks supported by the dataset.
Introduction
Perception in marine environments is notoriously difficult. Unlike terrestrial robotics, marine systems must contend with dynamic water surfaces, light refraction at the air-water interface, "caustics" (moving light patterns on the seabed), and varying turbidity. Most existing datasets are siloed, providing either aerial footage of the surface or underwater footage of the reef, but rarely both simultaneously. This lack of multi-modal data hinders the development of algorithms capable of tracking targets as they move between submerged and surfaced states, or creating stable maps of shallow areas where surface movement distorts aerial views. AMP2026 addresses this gap by providing a synchronized, multi-platform resource for marine perception.
Methodology
To capture the marine environment from multiple perspectives, researchers conducted field deployments in two distinct locations: the saltwater coral reefs of Barbados and a freshwater lake in Quebec. The data collection involved a fleet of diverse platforms:
- Aerial Drones: High-resolution 4K video captured at altitudes between 10m and 60m.
- Surface Vessels: Propeller-based vehicles equipped with downward-facing cameras.
- Submerged Robots: The "Aqua" robot, a hexapod submersible featuring synchronized stereo cameras.
- Divers: Human operators using handheld 4K action cameras to provide diver-perspective data.
A critical technical contribution of this dataset is the synchronization of telemetry and visual data across these platforms. For tracking, the team implemented a hardware solution involving GNSS antennas mounted on masts extending above the water from submerged robots, providing rare "ground truth" positional data for underwater targets.
Dataset Organization
The AMP2026 dataset is categorized into two primary research domains: Tracking and Mapping.
- Tracking Sequences: These include five classes of experiments, such as multi-drone tracking of submerged robots and single-diver tracking of underwater agents. These sequences feature scenarios both with and without GNSS ground truth.
- Mapping Sequences: These focus on repeated, geo-referenced observations of shallow-water regions, designed to test temporal fusion and distortion-compensation algorithms.
Key Insights and Research Applications
- The Air-Water Interface Challenge: The dataset highlights how refraction and surface glint act as high-frequency noise for traditional feature detectors. Multi-view data allows researchers to explore how multiple viewpoints can be fused to overcome these local distortions.
- Quantitative Subsurface Benchmarking: By providing GNSS data for submerged robots, the dataset enables quantitative error analysis for underwater tracking, moving beyond qualitative assessments.
- Environmental Contrast: The inclusion of both clear saltwater (Barbados) and turbid freshwater (Quebec) allows for testing the generalizability of perception models across varying optical conditions.
Results and Baseline Validation
The dataset comprises over 150 minutes of tracking footage and 30 distinct mapping sequences. It includes complex multi-target scenarios, such as multiple drones tracking two submerged robots simultaneously or drones tracking swimmers interacting with robots. The authors validated the dataset's utility by successfully applying YOLO-based object detection to underwater footage and generating high-quality orthomosaic maps of coral reefs using OpenDroneMap.
Conclusion and Future Work
While AMP2026 represents a significant advancement in marine robotics resources, future iterations aim to address current limitations. These include the addition of dense semantic segmentation labels (e.g., pixel-level labels for coral vs. sand) and expanding the variety of geographic locations. The team plans to incorporate more diverse deployments and establish standardized evaluation benchmarks to facilitate the comparison of tracking and mapping algorithms.
Dataset Access: AMP2026 on Hugging Face
Summary updated