Self-supervised Visual Representation Learning for Fine-Grained Ship Detection

The success of deep learning has greatly promoted the development of object detection algorithms. However, the performance of object detection algorithms in specific fields still needs to be improved. We focus on the field of fine-grained ship detection in our work. For fine-grained ship detection, labelling requires some professional knowledge, and it is difficult for the model to identify some less distinguishable categories. In this paper, a siamese network is proposed for self-supervised visual representation learning to reduce the reliance on a large amount of labelled data. Meanwhile, we collect and label a fine-grained ship dataset called HarborShips. We pretrain the backbone model in a self-supervised method to learn effective visual features by optimizing contrastive loss. The feature extractor is further fine-tuned for the downstream task of ship detection in a supervised way. In the task of ship detection, our model improves the mAP by 8.9% on our Harborships dataset.