Wilbur1240/MaSTr1325_512x384

MaSTr1325

A Maritime Semantic Segmentation Training Dataset for Small‐Sized Coastal USVs

---

Overview

MaSTr1325 (Maritime Semantic Segmentation Training Dataset) is a large-scale collection of real-world images captured by an unmanned surface vehicle (USV) over a two-year span in the Gulf of Koper (Slovenia). It was specifically designed to advance obstacle-detection and segmentation methods in small-sized coastal USVs. All frames are per-pixel annotated into three main semantic categories—obstacles/environment, water, and sky—and synchronized with IMU measurements from the onboard sensors.

• Total samples: 1 325
• Image resolution: 1 278 × 958 px (captured at 10 FPS via stereo USB-2.0 cameras mounted 0.7 m above water)
• Annotation categories (mask values):
  • Obstacles & Environment = 0
  • Water = 1
  • Sky = 2
  • Ignore/Unknown = 4

Because marine scenes often feature clear water/sky regions, a high-grade per-pixel annotation process (20 min/image) was used to ensure boundary accuracy. Any ambiguous “edge” pixels between semantic regions were marked as “ignore” (value 4), so they can easily be excluded from training/validation.

---

Authors & Citation

If you use MaSTr1325 in your work, please cite:

@inproceedings{bb_iros_2019,
  title     = {The MaSTr1325 dataset for training deep USV obstacle detection models},
  author    = {Bovcon, Borja and Muhovi{\v{c}}, Jon and Per{\v{s}}, Janez and Kristan, Matej},
  booktitle = {{2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}},
  year      = {2019},
  pages     = {3431--3438},
  organization = {IEEE}
}

@article{bb_ras_2018,
  title     = {Stereo obstacle detection for unmanned surface vehicles by IMU‐assisted semantic segmentation},
  author    = {Bovcon, Borja and Muhovi{\v{c}}, Jon and Per{\v{s}}, Janez and Kristan, Matej},
  journal   = {Robotics and Autonomous Systems},
  volume    = {104},
  pages     = {1--13},
  year      = {2018},
  publisher = {Elsevier}
}