Suzhen Fan ¹ Chengyang Song ² Haiyang Feng ¹ Zhibin Yu ^2*

• ¹ SANYA Oceanographic Institution, Ocean University of China, Sanya, China
• ² Ocean University of China, Qingdao, China

Fish re-identification(re-ID) is crucial for fish monitoring and can further promote aquaculture and fish breeding. Consequently, we have taken the first step in fish re-identification efforts. Synchronizing information from different cameras can accelerate or optimize reidentification performance. We constructed the first underwater fish re-identification benchmark dataset (FS48) under three camera conditions to promote the development of underwater reidentification. FS48 includes 48 different fish identities, 10,300 frames, and 39,088 bounding boxes, covering different lighting conditions during day and night and occluded and unoccluded background environments. We developed the first robust and accurate fish re-identification baseline, FSNet, which fuses information from three camera positions. FSNet extracts features from synchronized video frames from each camera position and fuses the synchronized information from the three positions. By combining information from three positions, FSNet achieves better re-identification performance. Our fish re-identification baseline helps improve overall re-test accuracy and evaluate the effectiveness of re-identification among detectors. The experimental results demonstrate that FS48 is universal and high-quality, and FSNet has an effective network design and good performance. Our dataset will be released upon acceptance of this paper.