a framework for underwater object detection
Due to underwater scenes are often accompanied by blur, scale variation, color shift, and texture distortion, generic object detection algorithms deployed on autonomous underwater vehicles will lead to poor performance. To address this problem, we propose an accurate and flexible underwater detection framework that employs spatial feature selection. Based on the systematic analysis, we observe that the center features of underwater creatures are more suitable for the classification task, while the edge features do well in the regression task. Moreover, due to underwater creatures tend to live in groups, causing the feature overlap, we also filtrate the features in the feature pyramid network. In addition, we also build an underwater detection dataset, denoted as UWD2021, which has more than 12,000 underwater images. Extensive experiments on URPC2018 and UWD2021 show that our method compares favorably with several state-of-the-art detectors.
[download link:]http://www.???.cn
We only provide the min-data for preview. For the whole dataset, please contact the author.
