Coralscapes: 2,075 High-Resolution Images for Coral Reef Segmentation

EPFL-ECEO released Coralscapes in 2025, providing 2,075 high-resolution images for dense semantic segmentation of coral reef environments. The dataset adopts the Cityscapes structure to facilitate benchmarking in the domain of underwater scene understanding.

Use Cases

• Training semantic segmentation models for pixel-level coral identification using the 1024x2048px resolution images
• Benchmarking underwater scene understanding algorithms against the dense semantic segmentation labels
• Ecological assessment of reef health by analyzing the distribution of pixel-level labels across the 2,075 images

Strengths

• 2,075 high-resolution images
• 1024x2048px resolution
• Standardized Cityscapes-style structure

Limitations

• Geographic bias towards US regions as indicated by metadata tags
• Domain-specific focus limited to coral reef ecosystems

Provenance

Source

EPFL-ECEO

Freshness

Last updated March 2025

Geography

United States