High-Resolution Underwater Creature Segmentation.

Area of Science:

• Computer Vision
• Marine Biology
• Robotics

Background:

• Underwater creature segmentation (UCS) is vital for marine research and robotics.
• Existing deep learning models struggle with high-resolution (HR) underwater imagery due to environmental distortions and biological complexities, leading to lost details and reduced precision.

Purpose of the Study:

• Introduce UCS4K, the first large-scale HR dataset for UCS, featuring 4,096 annotated images.
• Propose the Resolution-Asymmetric Dual-branch Alignment and Refinement (RADAR) network to overcome efficiency-receptiveness trade-offs in HR-UCS.

Main Methods:

• UCS4K provides 4x higher average resolution than prior datasets, covering diverse marine environments and species.
• The RADAR network employs a CNN branch for spatial details and a Transformer branch for global semantics on downsampled inputs.
• Global Semantic Alignment (GSA) and Bidirectional Collaborative Refinement (BCR) modules address inter-branch semantic misalignment and refine segmentation boundaries.

Main Results:

• RADAR achieves state-of-the-art performance on the UCS4K benchmark and existing datasets.
• The asymmetric network design efficiently captures long-range context without sacrificing spatial precision.

Conclusions:

• UCS4K establishes a new HR benchmark for underwater creature segmentation.
• The RADAR network offers a scalable and high-precision framework for HR-UCS, advancing marine research and robotics applications.