SCA-HNUST/mine-monorail-track-segmentation

Mine Monorail Track Segmentation Dataset

This dataset supports semantic segmentation of mine-use suspended monorails in simulated underground environments. It was collected in a dedicated experimental workshop provided by an industrial partner and includes images captured under diverse structural and illumination conditions to reflect realistic operational challenges.

📷 Data Collection

• Camera: Hikvision DS-2CD2820FWD
• Sensor: 1/2.8" Progressive Scan CMOS
• Resolution: 1280×720 (from native 1920×1080)

Images were captured in a controlled workshop that replicates key characteristics of underground monorail roadways, including track geometry, lighting variability, and common obstructions.

🗂️ Dataset Statistics

• Total images: 2,681
• Annotation tool: LabelMe (pixel-level precision)
• Classes:
  • 0: Background
  • 1: Monorail Track
• Track configurations:
  • Straight segments
  • Curved sections
  • Ascending slopes
  • Descending slopes
• Illumination conditions:
  • Normal lighting
  • Low-light
  • Overexposed

Each image is paired with a single-channel PNG mask where pixel values correspond to class IDs.

🔧 Data Augmentation (for reference)

To improve model robustness under harsh mining conditions, the authors employed a specialized data augmentation pipeline during training that simulates:

• Low-light scenarios through controlled brightness reduction,
• Sensor noise via Gaussian and salt-and-pepper noise injection,
• Dust interference using color overlays, texture synthesis, and visibility attenuation.

💡 Note: The raw dataset contains only original images and ground-truth masks—no augmented samples are included. Full technical details of the augmentation strategy are provided in the associated paper (Section 4.5).

📁 Directory Structure

├── images/
│ ├── Training_Input/
│ ├── Validation_Input/
│ └── Test_Input/
├── labels/
│ ├── Training_GroundTruth/
│ ├── Validation_GroundTruth/
│ └── Test_GroundTruth/
├── README.md
└── LICENSE