SUIM dataset

Dataset for semantic segmentation of underwater imagery. 

Paper: coming soon

Folder structure:

  • train_val/ contains 1525 paired samples for training/validation
    • images/: RGB images of underwater scenes
    • masks/: segmentation labels
      • Each RGB color represents a different object category
      • Further details in the paper (Section III)  
Object category RGB color code
Background (waterbody) 000 (black)
Human divers 001 (blue)
Aquatic plants and sea-grass 010 (green)
Wrecks and ruins 011 (sky)
Robots (AUVs/ROVs/instruments) 100 (red)
Reefs and invertebrates 101 (pink)
Fish and vertebrates 110 (yellow)
Sea-floor and rocks 111 (white)
  • TEST/ contains 110 paired samples for benchmark evaluation
    • images/: RGB test images
    • masks/: ground truths labels
      • Combined RGB masks are provided
      • Individual binary masks are also provided in separate folders