A wild fish image dataset for individual re-identification and phenotyping

Area of Science:

• Marine Biology
• Computer Vision
• Bioinformatics

Background:

• Automated wildlife monitoring requires large, diverse image datasets for accurate phenotyping and individual identification.
• Existing datasets often lack the longitudinal data needed to capture natural variation across years and individuals.

Purpose of the Study:

• To introduce Melops, a comprehensive longitudinal image dataset of corkwing wrasse (Symphodus melops).
• To provide a foundation for developing advanced computer vision techniques for wildlife research, including individual re-identification and phenotyping.

Main Methods:

• Collected 24,578 images of 9,861 individual corkwing wrasse over seven years using capture-mark-recapture.
• Utilized PIT-tagging for individual identification and standardized photography with detailed metadata (length, sex, reproductive state).
• Processed images into original and automatically cropped formats for deep learning applications.

Main Results:

• The Melops dataset contains extensive longitudinal data with thousands of individuals and repeated observations.
• Metadata includes precise measurements and biological information, enhancing its utility for ecological studies.
• The dataset supports various computer vision tasks such as re-identification, color pattern analysis, and sex classification.

Conclusions:

• Melops offers a valuable resource for advancing computer vision in wildlife monitoring, particularly for fish species.
• The dataset's longitudinal nature enables novel research into temporally aware re-identification and phenotypic changes in wild populations.
• Melops serves as a model for creating similar datasets for other taxa, promoting broader applications in biodiversity research.