🌊 MUOT-3M: A 3 Million Frame Multimodal Underwater Tracking Benchmark

March 9, 2026 · View on GitHub

Official repository for MUOT-3M
📄 MUOT-3M: A 3 Million Frames Multimodal Underwater Benchmark

🚀 Overview

MUOT-3M is the first large-scale multimodal underwater object tracking benchmark, designed to advance research in marine computer vision, underwater robotics, and multimodal learning.

Unlike previous RGB-only underwater datasets, MUOT-3M provides synchronized RGB, Enhanced RGB, Depth, and Language modalities, enabling robust learning under severe underwater degradations.

📊 Dataset Statistics

  • 🎥 3,030 underwater videos
  • 🖼 3.01 million frames
  • ⏱ 27.8 hours of footage
  • 🐠 677 fine-grained classes
  • 🌎 16 phyla
  • 🧬 124 families
  • 🏷 32 tracking attributes
  • 📦 Train/Test split: 70% / 30%
  • 🧠 Expert-validated annotations (marine biologist reviewed)

🌊 Modalities

Each sequence includes synchronized:

  • RGB frames\
  • Enhanced RGB frames\
  • Estimated depth maps\
  • Segmentation masks\
  • Language descriptions\
  • Bounding box annotations

The dataset captures real-world underwater challenges:

  • Low visibility\
  • Turbidity & backscatter\
  • Color attenuation\
  • Camouflage\
  • Swarm distractors\
  • Dynamic illumination\
  • Motion blur

📂 Dataset Access

The dataset is hosted on Hugging Face:

📈 Benchmark Comparison

MUOT-3M significantly exceeds existing underwater tracking datasets in:

  • Scale\
  • Class diversity\
  • Attribute coverage\
  • Multimodal design

🧪 Annotation Protocol

  • Semi-supervised bounding box generation\
  • Manual frame-by-frame verification\
  • Expert validation\
  • Segmentation mask refinement\
  • Ecological taxonomy validation

All sequences were curated to ensure:

  • Continuous target visibility\
  • Natural underwater scenes\
  • High annotation consistency

🎯 Applications

MUOT-3M supports research in:

  • Underwater object tracking\
  • Marine robotics\
  • Aquaculture monitoring\
  • Coral reef analysis\
  • Vision-language underwater models\
  • Multimodal representation learning

📜 Citation

@article{bakht2026muot_3m,
  title={MUOT\_3M: A 3 Million Frame Multimodal Underwater Benchmark and the 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

The dataset is released for research purposes under an open academic license.
Please check the Hugging Face page for full license details.

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