MVTD: Maritime Visual Tracking Dataset

Overview

MVTD (Maritime Visual Tracking Dataset) is a large-scale benchmark dataset designed specifically for single-object visual tracking (VOT) in maritime environments.
It addresses challenges unique to maritime scenes: such as water reflections, low-contrast objects, dynamic backgrounds, scale variation, and severe illumination changes—which are not adequately covered by generic tracking datasets.

The dataset contains 182 annotated video sequences with approximately 150,000 frames, spanning four maritime object categories:

• Boat
• Ship
• Sailboat
• Unmanned Surface Vehicle (USV)

MVTD is suitable for training, fine-tuning, and benchmarking visual object tracking algorithms under realistic maritime conditions.

---

Dataset Statistics

• Total sequences: 182
• Total annotated frames: 150,058
• Frame rate: 30 FPS and 60 FPS
• Resolution range:
  • Min: 1024 × 1024
  • Max: 1920 × 1440
• Average sequence length: ~824 frames
• Sequence length range: 82 – 4747 frames
• Object categories: 4

---

Dataset Structure

The dataset follows the GOT-10k single-object tracking format, enabling easy integration with existing tracking pipelines.

---

MVTD/
├── train/
│   ├── video1/
│   │   ├── frame0001.jpg
│   │   ├── frame0002.jpg
│   │   ├── ...
│   │   ├── groundtruth.txt
│   │   ├── absence.label
│   │   ├── cut_by_image.label
│   │   └── cover.label
│   ├── video2/
│   │   ├── frame0001.jpg
│   │   ├── frame0002.jpg
│   │   ├── ...
│   │   ├── groundtruth.txt
│   │   ├── absence.label
│   │   ├── cut_by_image.label
│   │   └── cover.label
│   └── ...
└── test/
    ├── video1/
    │   ├── frame0001.jpg
    │   ├── frame0002.jpg
    │   ├── ...
    │   └── groundtruth.txt
    ├── video2/
    │   ├── frame0001.jpg
    │   ├── frame0002.jpg
    │   ├── ...
    │   └── groundtruth.txt
    └── ...

---

Tracking Attributes

Each video sequence is categorized using nine tracking attributes:

1. Occlusion
2. Illumination Change
3. Scale Variation
4. Motion Blur
5. Variation in Appearance
6. Partial Visibility
7. Low Resolution
8. Background Clutter
9. Low-Contrast Objects

These attributes represent both maritime-specific and generic VOT challenges.

---

Data Collection

The dataset was collected using two complementary camera setups:

• Onshore static camera

• Large scale variations

• Perspective distortions

• Occlusions from vessels and structures

• Offshore dynamic camera mounted on a USV

• Strong illumination changes and glare

• Motion blur and vibrations

• Rapid viewpoint changes

This setup covers diverse maritime scenarios including:

• Coastal surveillance
• Harbor monitoring
• Open-sea vessel tracking

---

Evaluation Protocols

MVTD supports two evaluation settings.
For detailed implementation, evaluation scripts, and baseline tracker configurations, please visit the official GitHub repository:

🔗 https://github.com/AhsanBaidar/MVTD

Protocol I – Pretrained Evaluation

• Trackers pretrained on generic object tracking datasets
• Evaluated directly on the MVTD test split
• Measures generalization performance in maritime environments

Protocol II – Fine-Tuning Evaluation

• Trackers fine-tuned using the MVTD training split
• Evaluated on the MVTD test split
• Measures domain adaptation effectiveness for maritime tracking

Baseline Results

The dataset has been benchmarked using 14 state-of-the-art visual trackers, including Siamese, Transformer-based, and autoregressive models.
Results show significant performance degradation when using generic pretrained trackers and substantial gains after fine-tuning, highlighting the importance of maritime-specific data.

---

Intended Use

MVTD is suitable for:

• Single-object visual tracking
• Domain adaptation and transfer learning
• Maritime robotics and autonomous navigation
• Benchmarking tracking algorithms under maritime conditions

---

Citation

If you use this dataset, please cite:

@article{bakht2025mvtd,
  title={MVTD: A Benchmark Dataset for Maritime Visual Object Tracking},
  author={Bakht, Ahsan Baidar and Din, Muhayy Ud and Javed, Sajid and Hussain, Irfan},
  journal={arXiv preprint arXiv:2506.02866},
  year={2025}
}