Presenter: Levi Cai (Woods Hole Oceanographic Institution)

Description:
A major hindrance to the study of behavior of marine animals is the lack of availability of suitable datasets that capture their behavior in their natural habitats. While in recent years there has been substantial progress in autonomous visual object tracking, there has been less analysis to understand how well the current state-of-the-art applies specifically to animal tracking in the marine environment and what issues remain for their immediate deployment. To this end, we present an initial performance analysis of semi-supervised approaches to visual tracking of marine animals in order to establish a baseline for future work towards utilizing AUVs for in-situ and fully autonomous tracking. To accompany this study, we have assembled a novel dataset that attempts to capture different animal species, environments, and behaviors in the wild and is specific to marine animal tracking. The animals we include are octopuses, sharks, jellyfish, dolphins, and larvaceans. We focus on short-term visual tracker performance because we believe that for robust animal behavior data, continuous, un-interrupted tracks are preferred. Our results suggest that for some use-cases, such as tracking of solitary animals in the mid-water region, current approaches may be sufficient for immediate use. However, in other cases with more complex backgrounds and occlusions, these methods fail often and cannot be reliably used by a tracking robot without further improvements, for which we provide some insight.