Real-Time Background-Agnostic Fish Localization in Underwater Videos towards Autonomous Species Monitoring

This paper investigates a novel machine learning framework for autonomous, real-time fish localization in underwater videos with diverse backgrounds. The framework consists of three different algorithms from the family of deep learning and computer vision. Each of them is a good solution to one or more specific needs; however, each algorithm has its own limitations. Combining these methods using ensemble learning is a way to accomplish background-agnostic fish localization in real-time. A specific combination called weighted voting learns an optimal set of weights, such that the highest weight goes to the algorithm with the highest prediction accuracy. Results presented for two underwater datasets with significantly varying background and illumination demonstrate that weighted voting can produce consistent localization irrespective of the environment.