Knowledge Adaptation for Cross-Domain Moving Infrared Small Target Detection

Moving infrared small target detection has a wide range of applications. The existing methods heavily rely on training and test sets with a relatively uniform feature distribution. However, they suffer severe performance degradation when dealing with out-of-distribution data. To address this limitation, we propose a Knowledge adaptation-based method for Cross-domain moving infrAred small taRget Detection (KCARD), aiming to transfer source semantic information to the target domain by utilizing only a very small number of labeled target domain samples. Specifically, our method consists of two parts. The first is the knowledge adaptation branch. It uses the pre-trained Segment Anything Model (SAM) as a bridge. The labeled source domain image is used to generate a SAM box prompt for the target domain image, then SAM offers potential target position information for further detection. Another branch utilizes the proposed bidirectional deformable guided LSTM module to learn the spatio-temporal feature. In the end, the features of the two branches are fused to the detection head. Experimental results on public datasets show that our approach exhibits positive performance even with only 1% of the labeled target domain training samples. The code will be available upon acceptance.