meant learning how to communicate across disciplines and translate scientific requirements into computational tasks. The result is a dataset that integrates biological expertise directly into computer vision research. “We’re able to get something that others can’t,” he tells us. “Since we’re collaborating with marine biologists, this is the first dataset that has some domainspecific knowledge. We’re bringing new knowledge to the community so others can use it in their work. That’s something I'm proud of.” On that subject, the team hopes the dataset will enable new research directions in marine monitoring. David suggests that openvocabulary detection and instancelevel captioning remain particularly challenging problems for the field. Ziqiang highlights another direction: taxonomy. Identifying marine species often requires understanding hierarchical relationships between related organisms. “For species monitoring, there is some hierarchical taxonomy within the marine domain,” he explains. “That’s very different from recognizing common objects like cars, dogs, or cats.” Kit says the broader motivation for the work is the increasing reliance on AI systems to interpret the natural world. He recalls taking a photograph of a fish near the university’s pier and asking an AI assistant to identify it. He was sure it was a boxfish, but the system suggested it was a pufferfish. Later, a marine expert confirmed the AI tool was mistaken. “I was correct, actually, but I wasn’t certain because I’m not a marine biologist,” he recalls. “Imagine if a normal user just thinks this is a pufferfish because they trust the AI tool.” The experience reinforced his desire for deeper domain knowledge in AI systems. “It gave me more confidence to work harder to really improve things.” You can learn more about this work during Oral Session 3B: Image Recognition and Understanding I, Sunday 15:00–16:00 in AZ Ballroom 7, and during Poster Session 2, Sunday 16:00–17:45 in the Tucson Ballroom and Prefunction Space. 14 DAILY WACV Sunday Oral Presentation
RkJQdWJsaXNoZXIy NTc3NzU=