Who’s Who in the Anemone? A Benchmark for Clownfish Tracking in the Wild

Clownfish behavior and social structure offer interesting opportunities for marine biology research, however their analysis often requires time-consuming manual annotation of video data. Although computer vision can provide a means for automating this process, existing methods still struggle in long-range settings when fish move rapidly, appear visually similar, or frequently become occluded. To assess these challenges and motivate further research in this domain, we construct a clownfish tracking dataset for training and evaluation of multi-object tracking algorithms. We facilitate dataset construction by developing a semi-automated annotation tool, which we use to annotate 30,071 bounding boxes over 12,771 frames of clownfish video. This high-quality data allows us to benchmark and fine-tune several object detection and tracking pipelines, providing insight into their performance and failure modes within this challenging domain.