Single-stage object detection in the deep learning era: a critical survey of CNN-, transformer-, and hybrid architectures

Abstract Single-stage object detection has emerged as the dominant paradigm for real-time visual recognition, driven by demands from autonomous driving, unmanned aerial vehicles, and edge deployment. This survey provides a critical review of single-stage detectors developed between 2016 and 2026, structured around three architectural axes: (i) the convolutional family (YOLOv1–v26, SSD, RetinaNet, FCOS, CornerNet/CenterNet, EfficientDet, RTMDet), (ii) the Transformer family (DETR and its descendants: Deformable DETR, DAB-DETR, DN-DETR, DINO, RT-DETR), and (iii) hybrid CNN–Transformer designs (DEYO, RT-DETR). Rather than cataloging models chronologically, we organize the analysis around three contested design questions in contemporary literature: anchor-based versus anchor-free regression, NMS-based versus NMS-free inference, and dense pixel-level versus sparse query-based prediction. We synthesize reported performance for 35 representative models on MS COCO into a literature-based meta-analysis, explicitly documenting input resolution, hardware, training schedule, and the primary source for each row, so readers can interpret comparisons fairly. We further provide a comparison with fourteen existing surveys to clarify this paper's incremental contribution, discuss methodological reporting inconsistencies in the primary literature, and outline a reproducibility checklist for future single-stage detection research. We close with a problem-driven discussion of application domains (small-object detection, UAV imagery, traffic monitoring, agriculture) and research directions, including NMS-free training, open-vocabulary detection, foundation-model-assisted detection, and CNN–Transformer hybridization.