Underwater Object Detection Under Domain Shift

Abstract

There is increasing interest in using deep learning–based object recognition algorithms to automate the labeling of image data collected from marine surveys. However, underwater object detection is a particularly challenging problem due to changes in scattering and absorption of light, and spotty data collection efforts, which rarely capture the broad variability. Using deep learning–based object detection systems for long-term or multisite marine surveying is further complicated by shifting data distributions between training and testing stages. Using data from the 100 Island Challenge, we investigate how object detection performance is impacted by changes in site characteristics and imaging conditions. We demonstrate that the combined use of data augmentation and unsupervised domain adaptation techniques can mitigate performance drops in the presence of domain shift. The proposed methodologies are broadly applicable to observational data sets in marine and terrestrial environments where a single algorithm needs to adapt to and perform comparably across changing conditions.