Deep Learning-based Detection of Military Equipment with Synthetic Infrared Images

Recent advances in deep learning-based object detection have highlighted the potential for applications in military domains. However, the scarcity of infrared(IR) training data remains a major limitation. In this work, we compare to representative EO-to-IR generation approaches and evaluate both image synthesis quality and downstream detection performance. Experimental results show that preserving spatial consistency with EO annotations, which means ensuring that EO-derived annotations remain valid by preventing hallucinated or misplaced objects in the generated IR images, is more important for object detection than photorealistic image generation. These findings suggest that our approach offers a practical training strategy for military IR object detection when real IR data are scarce, particularly by preserving spatial consistency between EO images and generated IR images.