A realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis.

Area of Science:

• Computer Vision
• Marine Biology
• Sustainable Fisheries

Background:

• Visual analysis of fish habitats is crucial for fisheries management and environmental conservation.
• Existing deep learning datasets for fish analysis lack the complexity of natural underwater environments.
• There is a need for comprehensive benchmarks to advance underwater computer vision for ecological monitoring.

Purpose of the Study:

• Introduce DeepFish, a large-scale benchmark dataset for underwater fish analysis.
• Facilitate training and testing of computer vision models for tasks beyond simple classification.
• Enable automated monitoring of fish populations, including counting, localization, and size estimation.

Main Methods:

• Collected approximately 40,000 underwater images from 20 distinct marine habitats in tropical Australia.
• Augmented initial classification labels with point-level and segmentation data for richer annotations.
• Evaluated the performance of state-of-the-art computer vision approaches on the DeepFish benchmark.

Main Results:

• The DeepFish dataset provides a comprehensive resource for diverse underwater fish analysis tasks.
• Models pre-trained on ImageNet demonstrate capability but indicate significant room for improvement.
• The benchmark highlights the challenges and potential of deep learning in complex underwater scenes.

Conclusions:

• DeepFish serves as a valuable testbed for developing advanced underwater computer vision techniques.
• Further research is motivated to enhance model performance in complex marine environments.
• This benchmark supports the development of sustainable fisheries through improved ecological monitoring.