Refined Datasets and Saliency Map Analysis for Underwater Object Detection

Underwater operations performed by divers are inefficient and dangerous. This encourages the development of underwater drones to replace human operations. Underwater object detection technology, in particular, plays a role in estimating the current position, controlling, and grabbing in the underwater drone. In this paper, we propose underwater object detection algorithms using refined underwater datasets to investigate model selection and hyperparameters. From the simulation results, the optimal parameter set for YOLOX is determined and a saliency map generation method is integrated into the YOLOX framework for the validity of the reasoning basis.