Mcships: A Large-Scale Ship Dataset For Detection And Fine-Grained Categorization In The Wild

This paper introduces a multi-category ship dataset (called McShips), which is a challenging and large-scale dataset aimed at ship detection and fine-grained categorization. The McShips dataset includes 14,709 annotated images of ships belonging to 6 classes of warships and 7 classes of civilian ships. Each image is carefully annotated with a bounding box and ship class label. The dataset is challenging due to the following two reasons: First, there is little inter-class variation as ships have very similar ship shapes; Second, there is very large intra-class variation since the ships within the same class may be significantly different due to viewpoint variations, weather condition variations, illumination variations, scale changes, occlusion, cluttered background and so on. Based on the McShips dataset, we present the detection and fine-grained categorization performance of three baseline detectors on our dataset, and make a comparison to identify the strengths and weaknesses of the baseline detection algorithms. We hope the presented McShips dataset would advance research and applications on ship detection and finegrained categorization.