Researchers have conducted a comparative study of out-of-the-box technology for automatic target detection and recognition (ATD/R). They benchmarked several state-of-the-art models, including YOLO series and DETR framework variations, on a newly acquired military relevant dataset. The study found that bigger models perform better than smaller ones, and DETR-based models show promising results compared to the YOLO series. However, all models struggle with detecting small objects in an Air-to-Ground (A2G) scenario, even after fine-tuning on an out-of-domain A2G dataset. The study concludes that in-domain training is still crucial for creating capable ATD/R systems. --- Why it matters: This study matters to researchers and engineers working on AI applications for military decision support and autonomous operations because it highlights the limitations of using publicly available models and datasets, and emphasizes the importance of domain-specific training for achieving high performance. Source: https://arxiv.org/abs/2608.17917