AhsanBB/MUOT_3M-A_3_Million_Frame_Underwater_Object_Tracking_Dataset

🌊 MUOT-3M: The Largest Multimodal Underwater Object Tracking Dataset

Official repository for MUOT-3M
📄 MUOT-3M: The Largest Multimodal Underwater Object Tracking Dataset and MUTrack Tracking Method

---

🚀 Overview

MUOT-3M is currently the largest underwater object tracking dataset, containing over 3 million annotated frames across 3,030 underwater videos with synchronized multimodal annotations.

The benchmark is designed to advance research in:

• Underwater object tracking
• Marine computer vision
• Underwater robotics
• Vision-language learning
• Multimodal representation learning

Unlike previous underwater tracking datasets that rely only on RGB imagery, MUOT-3M introduces synchronized:

• RGB
• Enhanced RGB
• Depth
• Segmentation
• Language modalities

This enables robust learning under severe underwater degradations and challenging marine environments.

---

🌟 Why MUOT-3M?

Existing underwater tracking benchmarks suffer from several limitations:

• Limited dataset scale
• RGB-only annotations
• Low class diversity
• Limited ecological realism
• Sparse attribute annotations
• Lack of multimodal synchronization

MUOT-3M addresses these challenges by introducing the first large-scale multimodal underwater tracking benchmark with:

✅ Over 3 million frames
✅ 3,030 underwater videos
✅ 677 fine-grained marine classes
✅ Synchronized RGB, depth, segmentation, and language modalities
✅ Expert-validated annotations
✅ Real-world underwater degradations and ecological diversity

---

📊 Dataset Statistics

Property	Value
🎥 Videos	3,030
🖼 Frames	3.01 Million
⏱ Footage Duration	27.8 Hours
🐠 Fine-Grained Classes	677
🌎 Phyla	16
🧬 Families	124
🏷 Tracking Attributes	32
📦 Train/Test Split	70% / 30%
🧠 Annotation Validation	Marine Biologist Reviewed

---

🌊 Modalities

Each sequence includes synchronized multimodal annotations:

• RGB frames
• Enhanced RGB frames
• Estimated depth maps
• Segmentation masks
• Natural language descriptions
• Bounding box annotations

---

🌊 Underwater Challenges

MUOT-3M captures real-world underwater tracking challenges including:

• Low visibility
• Turbidity and backscatter
• Color attenuation
• Motion blur
• Dynamic illumination
• Camouflage
• Swarm distractors
• Fast target deformation
• Occlusion
• Scale variation

---

📈 Benchmark Comparison

MUOT-3M significantly exceeds previous underwater tracking datasets in scale, diversity, and multimodal design.

Dataset	Frames	Videos	Modalities
DeepSea MOT	2K	4	RGB
MFT25	48K	15	RGB
UVOT400	275K	400	RGB
WebUOT-1M	1.1M	1,500	RGB + Language
MUOT-3M	3.01M	3,030	RGB + Enhanced RGB + Depth + Segmentation + Language

---

🧪 Annotation Protocol

MUOT-3M follows a rigorous multi-stage annotation pipeline:

Annotation Process

1. Semi-supervised bounding box generation
2. Manual frame-by-frame verification
3. Segmentation mask refinement
4. Language annotation generation
5. Expert marine biology validation

Quality Assurance

All sequences were curated to ensure:

• Continuous target visibility
• Annotation consistency
• Ecological correctness
• Natural underwater scenes
• Robust multimodal synchronization

---

📂 Dataset Structure

MUOT-3M/
│
├── RGB/
├── Enhanced_RGB/
├── Depth/
├── Segmentation/
├── Language/
├── Annotations/
│   ├── train/
│   └── test/
│
└── metadata/

---

🤗 Dataset Access

The dataset is hosted on Hugging Face.

🔹 Full Dataset

👉 https://huggingface.co/datasets/AhsanBB/MUOT_3M-A_3_Million_Frame_Underwater_Object_Tracking_Dataset

---

🔍 Keywords

underwater object tracking, underwater tracking benchmark, multimodal tracking, marine computer vision, underwater robotics, RGB-D tracking, underwater MOT, vision-language tracking, underwater dataset, multimodal underwater benchmark, marine AI, segmentation, depth estimation

---

🎯 Applications

MUOT-3M supports research in:

• Underwater object tracking
• Marine robotics
• Autonomous underwater vehicles (AUVs)
• Aquaculture monitoring
• Coral reef analysis
• Ecological monitoring
• Vision-language underwater models
• Multimodal representation learning
• Underwater segmentation
• Marine biodiversity analysis

---

📄 Paper

MUOT-3M: The Largest Multimodal Underwater Object Tracking Dataset and MUTrack Tracking Method

📄 arXiv: https://arxiv.org/abs/2602.18006

---

📜 Citation

@article{bakht2026muot_3m,
  title={MUOT-3M: The Largest Multimodal Underwater Object Tracking Dataset and MUTrack Tracking Method},
  author={Bakht, Ahsan Baidar and Alansari, Mohamad and Din, Muhayy Ud and Naseer, Muzammal and Javed, Sajid and Hussain, Irfan and Matas, Jiri and Mahmood, Arif},
  journal={arXiv preprint arXiv:2602.18006},
  year={2026}
}

---

📜 License

MUOT-3M is released for academic and research purposes under an open research license.

Please check the Hugging Face dataset page for full licensing details.

---

🙏 Acknowledgements

We thank:

• Marine biology experts for annotation validation
• Underwater robotics researchers
• The computer vision community
• Contributors and collaborators supporting underwater AI research

---

🌊 MUOT-3M establishes a new foundation for scalable multimodal underwater object tracking research.