MMOS: A Multi-Sensor UAV Dataset for Maritime Object Detection and Surveillance

MMOS is a multi-sensor unmanned aerial vehicle (UAV) dataset designed for research in maritime object detection, surveillance, and search and rescue (SAR). The dataset was collected using a DJI Matrice 30T under real-world coastal operating conditions and provides synchronized recordings from multiple sensing streams with different fields of view and sensing characteristics.

The dataset contains synchronized Screen, Wide, Zoom, and Thermal streams, together with frame metadata and UAV/gimbal telemetry. This configuration enables research on object detection under different sensing conditions, object-scale analysis, multi-sensor perception, sensor fusion, and geo-referenced maritime perception.

Dataset version: 1.0
License: CC BY 4.0
Paper status: Submitted to IEEE Access; publication details will be added after acceptance.

---

Dataset Overview

The MMOS dataset was collected over five recording sessions, resulting in approximately 46 minutes of video data distributed across four synchronized streams per session.

The four streams are:

• Screen — recording of the operator's control interface during the flight.
• Wide — wide-angle visible-spectrum imagery providing a broad field of view.
• Zoom — visible-spectrum imagery acquired using the dedicated telephoto/zoom sensor.
• Thermal — long-wave infrared (LWIR) imagery.

The Screen and Zoom streams can contain visually similar observations because the DJI Matrice 30T controller uses a sensor-switching mechanism to provide a continuous zooming experience. At lower magnification levels, the displayed view is provided by the Wide camera, while the dedicated Zoom sensor becomes active at higher magnification levels. The Zoom stream therefore starts from its own optical baseline rather than from the Wide camera's 1× view.

In addition to the imagery, the dataset provides synchronized telemetry and frame-level metadata.

---

Dataset Structure

The repository is organized as follows:

MMOS/
├── annotations/
│   └── v1.0.zip
│
├── frames/
│   ├── MMOS_seq01_frames.zip
│   ├── MMOS_seq02_frames.zip
│   ├── MMOS_seq03_frames.zip
│   ├── MMOS_seq04_frames.zip
│   └── MMOS_seq05_frames.zip
│
├── metadata/
│   ├── seq01_frames.csv
│   ├── seq01_hover_segments.csv
│   ├── seq02_frames.csv
│   ├── seq02_hover_segments.csv
│   ├── seq03_frames.csv
│   ├── seq03_hover_segments.csv
│   ├── seq04_frames.csv
│   ├── seq04_hover_segments.csv
│   ├── seq05_frames.csv
│   └── seq05_hover_segments.csv
│
├── mmos_benchmark/
│   └── benchmark_v1.0.zip
│
├── telemetry/
│   └── telemetry.zip
│
└── videos/
    ├── MMOS_seq01_videos.zip
    ├── MMOS_seq02_videos.zip
    ├── MMOS_seq03_videos.zip
    └── MMOS_seq04_videos.zip

annotations/

Contains the manually reviewed object annotations.

v1.0.zip corresponds to the annotation release used to obtain the experimental results reported in the associated paper.

The initial annotation protocol defines four object categories:

• Vessel
• Human
• Buoy
• Debris

The annotations are provided in YOLO format.

Future annotation improvements or additional annotations may be released as subsequent versions. Versioned releases are maintained so that previously reported experiments remain reproducible.

frames/

Contains the extracted image frames organized by recording sequence.

Each sequence contains the available sensing streams corresponding to that flight session.

The frame release is based on the synchronized flight logs, with frames selected at approximately 10 Hz (10 FPS) to match the available flight-recording metadata.

metadata/

Contains frame-level metadata and manually identified hover segments.

The *_frames.csv files provide information associated with the extracted frames.

The *_hover_segments.csv files identify segments in which the UAV remained approximately stationary for extended periods. These files can be used to reproduce or modify the preprocessing performed for the benchmark experiments.

mmos_benchmark/

Contains the benchmark configuration and resources associated with the experiments reported in the paper.

benchmark_v1.0.zip corresponds to the benchmark based on annotation release v1.0.

The benchmark is intended to facilitate reproducibility of the reported object detection experiments.

telemetry/

Contains telemetry information synchronized with the recorded video streams.

The available telemetry fields include:

• filename
• video_filename
• time
• datetime
• latitude
• longitude
• ground_level_altitude
• sea_level_altitude
• gimbal_pitch
• gimbal_roll
• gimbal_yaw

The telemetry is sampled at approximately 10 Hz (100 ms intervals).

These data can support research involving UAV motion, camera orientation, spatial analysis, and geo-referenced perception.

videos/

Contains the original recorded video streams organized by recording sequence.

The video files provide access to the source data from which the released frame datasets were extracted.

---

Data Splits

For the benchmark experiments, the dataset is divided at the sequence level to avoid temporal leakage between training, validation, and test data.

The split used for the reported experiments is:

Split	Sequences
Training	seq01, seq03
Validation	seq04, seq05
Test	seq02

Frames from the same recording sequence are therefore not distributed across different subsets.

This strategy prevents temporally adjacent frames and highly similar observations from appearing simultaneously in training and evaluation sets.

---

Object Classes

The initial annotation release contains four semantic object categories:

Class	Description
Vessel	Boats and ships, regardless of their apparent size in the image
Human	Individuals visible in the water or near the shoreline
Buoy	Floating markers and navigation buoys
Debris	Floating marine debris and other identifiable waste materials

For the benchmark experiments, these original categories were additionally grouped into two broader semantic categories:

• Vessels — Vessel
• Surface Objects — Human, Buoy, and Debris

Object scale is treated independently from semantic class. Small, medium, and large objects are determined from bounding-box area according to the COCO object-size definitions.

---

Annotation Versioning

The experimental results reported in the associated paper were obtained using annotation release v1.0.

The annotation repository may be updated after the initial release to improve existing bounding boxes and/or add previously unannotated objects. Such changes will be distributed as new versioned releases rather than replacing the annotations used for the reported experiments.

This versioning scheme is intended to preserve the traceability and reproducibility of published benchmark results.

---

Benchmark

The MMOS benchmark evaluates three representative object detection paradigms:

• YOLOv8 — one-stage detection
• Faster R-CNN — two-stage detection
• RT-DETRv2 — transformer-based detection

The benchmark considers the different sensing modalities and evaluates detection performance across semantic categories and object scales.

The benchmark experiments use images resized to 640 × 640 pixels.

The released benchmark corresponds to version 1.0 of the annotations.

---

Recommended Citation

If you use MMOS in your research, please cite the associated dataset paper:

@dataset{barreto_mmos,
  author    = {Barreto, Joniel B. and da Silva, Juliana M. F. and
               Caetano, Filipe F. and Gehlen, Alexsandro and
               Morais, Edmilson S. and Duek, Carlos and
               Verri, Filipe A. N. and Forster, Carlos H. Q.},
  title     = {MMOS: A Multi-Sensor UAV Dataset for Maritime Object Detection and Surveillance},
  year      = {2026},
  publisher = {Hugging Face},
  note      = {Dataset associated with a manuscript submitted to IEEE Access}
}

Once the associated paper is accepted and published, the citation will be updated with the final bibliographic information and DOI.

---

License

MMOS is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to:

• Share — copy and redistribute the material in any medium or format.
• Adapt — remix, transform, and build upon the material for any purpose, including commercially.

Under the following condition:

• Attribution — appropriate credit must be given, a link to the license must be provided, and any changes must be indicated.

Full license information is available at:

https://creativecommons.org/licenses/by/4.0/

When using the dataset, please cite the associated dataset paper as indicated above.

---

Disclaimer

The dataset was collected under real-world coastal operating conditions and may contain variations in illumination, viewpoint, object scale, background appearance, and sensor characteristics.

The annotations represent the annotation release identified in this repository and may be improved in subsequent versions. Users should cite the specific dataset and annotation version used in their experiments.

---

Contact

For questions regarding the dataset, annotation releases, or benchmark experiments, please refer to the associated paper or contact the authors through their institutional affiliations.