Vessel instance mining for vessel detection with limited annotations

Recently, video surveillance-based vessel detection has garnered significant attention. However, current detectors primarily rely on high-quality densely annotated datasets with extensive annotations, incurring high costs that limit real-world deployment. To address this, this paper proposes VeIM, a novel vessel instance mining framework designed for extremely limited annotation settings. Specifically, a Potential Vessel Instance Recall module is introduced to retrieve latent unannotated instances, enhancing training efficiency. Furthermore, a Pseudo Vessel Instance Discrimination module is proposed, which validates the mined instances by leveraging intra-class consistency and inter-class discrimination. Extensive experiments on the SeaShip dataset demonstrate that VeIM significantly outperforms baseline methods and achieves performance comparable to Fully Supervised Object Detection, despite using only a fraction of the annotations.