Dual Consistency Learning Network for Multimodal Semisupervised Oriented Ship Detection

Accurate oriented ship detection is critical for maritime surveillance applications. Leveraging the complementary strengths of different image modalities can improve detection performance. However, multimodal inconsistencies and the dependence on costly, time-consuming paired annotations significantly limit the scalability of multimodal methods. While existing semi-supervised methods provide solutions to annotation scarcity, most of these methods struggle to generate reliable pseudo labels specifically for ship orientation estimation, particularly under challenging environmental conditions and for densely arranged ships. To this end, we pioneer the Multimodal Semi-supervised Oriented Object Detection (MSOOD) task and propose the Dual Consistency Learning Network (DCLNet) framework, which integrates: 1) The Visual Knowledge Consistency Learning (VKCL) module aligns visible-infrared pseudo labels by minimizing the inter-modal consistency loss and introduces a Large Visual Language Model (LVLM) to evaluate the rationality of pseudo labels and correct their confidence. 2) The Spatial Consistency Learning (SCL) module employs the Mask-augmented Transformer (MAT) for hierarchical prediction of the position, angle, and size of the rotated bounding box with three spatial consistency losses. 3) The Multimodal Interaction and Fusion (MIF) module introduces Efficient 2D Scanning (ES2D) layer and Cross-Modality Fusion Attention (CFA) for complementary feature fusion. Experimental results demonstrate that our proposed approach significantly outperforms state-of-the-art models on the constructed DOTAv1.5-Ships* and FAIR1M-v2.0-Ships* remote sensing datasets.