11 DAILY WACV Sunday Yuk Kwan (David) Wong coverage and annotation detail, the dataset aims to reflect the diversity encountered in real-world marine surveys more accurately. The second challenge relates to how computer vision tasks are typically defined. Many benchmarks rely on simple tasks such as image classification or short captions. For marine scientists, however, those outputs are often too limited to be useful in practice. Marine ecosystems are complex, and researchers often need richer descriptions of the organisms they observe. “Using a single-label task like image classification – this is an image of a fish, this is an image of a dolphin – may not be very helpful,” David points out. “They need something more professional.” To better match those needs, the team designed a set of tasks tailored to marine research. The dataset supports object detection, instancelevel captioning, and visual grounding, allowing models to both identify organisms and generate detailed descriptions of them. Crucially, many of the captions describe morphological features, such as body shape, color patterns, or fin structures, that marine biologists use to distinguish between similar species. Building those annotations required expertise beyond computer science. The project involved close collaboration between computer vision researchers and marine biologists. “In the author list, you’ll notice there are some people from computer science, and there are biologists as well,” David says. “It’s a multi-disciplinary paper.” Marine experts helped annotate images and verify descriptions to ensure that the dataset reflects real biological knowledge. For Ziqiang, another important aspect of the work is enabling models to recognize species that may not already appear in training data. “Nowadays, for underwater
RkJQdWJsaXNoZXIy NTc3NzU=