EUWOD-16: An Extended Dataset for Underwater Object Detection

Remotely Operated Vehicles (ROVs) along with vision based underwater object detection techniques can assist underwater exploration and research by identifying specific objects, such as shipwrecks, marine life, and man-made debris. These object detection algorithms require large datasets for training. Since there are very few datasets available for underwater objects, in this paper, an Extended Underwater Object Detection dataset with 16 object categories (EUWOD-16) was constructed. This was achieved by building a new annotated dataset consisting of divers, artifacts and various marine species, and merging it with existing underwater object detection datasets by redefining their annotations. Later, the dataset was evaluated by a modified YOLOv5n architecture with GhostNet. This method involved the selective additions of ghost blocks in appropriate places to decrease the number of network parameters and FLOPs without a significant decrease in performance, and a Bi-FPN connection for a refined feature fusion pathway. The proposed model achieved higher accuracy at a comparatively lower number of parameters (and FLOPs) than both YOLOv5n and GhostNet.