Underwater object detection using collaborative weakly supervision

Despite recent progress in deep learning, underwater object detection remains a challenge where noisy and imprecise images are provided as sources of supervision. This paper presents a novel underwater detection approach in the framework of weakly supervised learning. The idea is to train two deep learning detectors simultaneously, and let them teach each other based on the selection of the cleaner samples they see during the training. The backbone of each detector is Yolov5, which achieves balance between accuracy and speed. The method is tested on the URPC2021 benchmark dataset, and achieves state-of-the-art performance. Compared with the original Yolov5, the dual training mechanism improves the recognition accuracy by 10%.