An underwater image dataset for occlusion-aware fish instance segmentation.

Area of Science:

• Aquaculture technology
• Computer vision
• Machine learning

Background:

• Accurate underwater fish perception is crucial for intelligent aquaculture.
• Fish occlusion in images hinders precise instance segmentation.
• Large-scale datasets with detailed annotations are needed for robust models.

Purpose of the Study:

• Introduce the Fish Occlusion Dataset (FOD) for occlusion-aware fish instance segmentation.
• Provide a benchmark for evaluating instance segmentation models under various occlusion levels.
• Facilitate advancements in automated fish monitoring and management.

Main Methods:

• Collected 14,376 underwater images from a precision aquaculture base.
• Annotated 144,894 fish instances with pixel-level detail across three occlusion levels (whole, part, fragment).
• Included both original and synthesized images for comprehensive model training and evaluation.

Main Results:

• The Fish Occlusion Dataset (FOD) enables fine-grained performance assessment across occlusion scenarios.
• Eight instance segmentation models were benchmarked, evaluating detection-based and proposal-free architectures.
• Mask2Former demonstrated superior overall segmentation performance, particularly in highly occluded conditions.

Conclusions:

• The FOD is a valuable resource for developing and evaluating fish instance segmentation models.
• Addressing occlusion is key to improving fish perception in intelligent aquaculture.
• Advanced models like Mask2Former show promise for real-world applications in aquaculture.